JP5364578B2 - Method and system for transductive data classification and data classification method using machine learning technique - Google Patents

Abstract

A system, method, data processing apparatus, and article of manufacture are provided for classifying data. Data classification methods using machine learning techniques are also disclosed.

Description

  The present invention relates generally to a method and apparatus for data classification. More particularly, the present invention provides an improved transductive machine learning method. The invention also relates to a novel application using machine learning techniques.

  Data processing methods are gaining importance in the information age, and more recently, all areas of life, including scanned documents, web materials, search engine data, text data, images, audio data files, among others. With the rapid increase of electronic data in Japan, its importance is increasing.

  One area that has just begun exploration is the non-manual classification of data. In many taxonomies, machines or computers must learn based on manually entered and generated rule sets and / or manually generated training examples. In machine learning in which training examples are used, the number of learning examples is generally smaller than the number of parameters that need to be estimated. That is, there are many solutions that satisfy the constraints given by the training examples. The challenge of machine learning is to find a solution that is sufficiently generalized despite the lack of constraints. Accordingly, there is a need to overcome these and / or other problems associated with the prior art.

  What is further needed is a practical application for all kinds of machine learning techniques.

  In a computer-based system, the data classification method according to one embodiment of the present invention is a training example of data points where labeled data points should be included in the specified category or excluded from the specified category Receiving labeled data points each having at least one label indicating whether it is a data point training example; receiving unlabeled data points; and at least one of labeled and unlabeled data points Receiving a predetermined cost factor, at least one cost factor, a labeled data point, and an unlabeled data point as training examples, and iteratively calculating and transducing using maximum entropy identification (MED). Training the active classifier for each iteration of the computation. The cost factor for unlabeled data points is adjusted as a function of the expected label value, the prior probability of the data point label is adjusted based on the estimation of the class membership probability of the data point, and a trained classifier is applied Classifying at least one of the unlabeled data point, the labeled data point, and the input data point and classifying the classified data point, or a derivative thereof, to the user, another system, and Outputting to at least one of the other processes.

  A method for classifying data according to another embodiment of the present invention includes providing computer executable program code to be deployed and executed on a computer. The program code includes at least one label indicating whether a data point is an example of training a data point to be included in a specified category or an example of training a data point excluded from a specified category, respectively. Instructions to access stored labeled data points in computer memory, instructions to access unlabeled data points from computer memory, and at least labeled and unlabeled data points from computer memory For each instruction that accesses a given cost factor and each iteration of the calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point label is the class membership probability of the data point At least one stored cost adjusted according to an estimate of Instructions to train a maximum entropy identification (MED) transductive classifier by iterative computation using factors, stored labeled data points, and unlabeled data points as training examples, unlabeled data points, labeled Instructions for applying a trained classifier to classify at least one of the data points and the input data points and the classification of the classified data points, or a derivative thereof, to a user, another system, And instructions for outputting to at least one of the other processes.

  A data processing apparatus according to another embodiment of the present invention is (i) a training example for data points that are included in a specified category, or for data points that are excluded from the specified category. Labeled data points each having at least one label indicating whether it is a training example; (ii) unlabeled data points; and (iii) at least one predetermined cost factor for labeled and unlabeled data points; , And at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples, using a transductive maximum entropy identification (MED) ) To repeatedly teach a transductive classifier And a Restorative classifier training devices. At each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, the prior probability of the data point label is adjusted according to the estimate of the class membership probability of the data point, and the transductive A classifier trained by a classifier training device is used to classify and classify at least one of unlabeled data points, labeled data points, and input data points, and , Output to at least one of another system, another step.

  A product according to another embodiment of the present invention comprises a computer readable program storage medium, which is an example of training for a data point or a designated data point to be included in a designated category. Receiving labeled data points each having at least one label indicating whether it is a training example for data points excluded from the category; receiving unlabeled data points; and unlabeled data points and labels Receiving at least one predetermined cost factor for the none data point and at each iteration of the MED calculation, the cost factor for the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point is the class of the data point At least one stored cost factor that is adjusted according to the attribution probability estimate Training the transductive classifier with iterative maximum entropy identification (MED) using the stored labeled data points and the stored unlabeled data points, and applying the trained classifier; Classifying at least one of unlabeled data points, labeled data points, and input data points, and classifying the classified data points, or a derivative thereof, to a user, another system, and another step And unambiguously embodying one or more programs comprising computer-executable instructions for performing a taxonomy, including outputting to at least one of the two.

In a computer-based system, the method for classifying unlabeled data according to another embodiment of the present invention is an example of training for data points whose data points should be included in a specified category or excluded from a specified category. Receiving labeled data points each having at least one label indicating whether it is a training example for a data point being labeled, receiving labeled and unlabeled data points, labeled data points and labels Receiving pre-label probabilities for unlabeled data points; receiving at least one predetermined cost factor for labeled data points and unlabeled data points; and each labeled and labeled based on prior probabilities of data point labels None Step to determine the expected label for a data point If, until the data value is substantially converges, the following substeps, namely,
Generating a scaled cost value for each unlabeled data point in proportion to the absolute value of the expected label of the data point;
Decisions that use labeled and unlabeled data as training examples according to their expected labels and minimize KL divergence for prior probability distributions of decision function parameters given included training examples and excluded training examples Training a classifier by determining a function;
Using a trained classifier to determine classification scores for labeled and unlabeled data points;
Calibrating the output of the trained classifier to class membership probabilities;
Updating the prior probabilities of labels of unlabeled data points according to the determined class membership probabilities;
Determining the probability distribution of labels and margins using maximum entropy identification (MED), using the updated prior probabilities and previously determined classification scores;
Using the previously determined label probability distribution to calculate a new expected label;
Updating the expected label for each data point by incorporating a new expected label along with the expected label obtained from the previous iteration;
Repeating the steps.
The classification of input data points, or a derivative thereof, is output to at least one of a user, another system, and another process.

  A method for classifying documents according to another embodiment of the present invention includes receiving at least one labeled seed document having a known confidence level for label assignment, receiving an unlabeled document, and at least one predetermined Receiving a cost factor and transductive by iterative calculation using at least one predetermined cost factor adjusted as a function of expected label value for each iteration of the calculation, at least one seed document, and an unlabeled document Training a classifier, storing a confidence score for an unlabeled document after at least some iterations, and identifying an identifier for an unlabeled document with the highest confidence score for a user, another system, and another process Output to at least one of the Tsu including and up, the.

  According to another embodiment of the present invention, a method for analyzing a document associated with a legal discovery procedure includes receiving a document associated with a legal matter, and performing a document classification technique on the document. Outputting an identifier of at least a part of the document based on the classification of the document.

  A method for organizing data according to another embodiment of the present invention includes receiving a plurality of labeled data items and, for each of the plurality of categories, selecting a subset of the data items for each of the plurality of categories. Setting uncertainties for data items in each subset to substantially zero; setting uncertainties for data items not in the subset to a predetermined value that is not substantially zero; and uncertainties, subsets Training the transductive classifier by iterative computation using the data items in and those that are not in the subset as training examples, and label the trained classifier to classify each of the data items The steps to apply to each of the data items, and the classification or derivation of the input data items And it includes user, and outputting another system, and at least one of another process, the.

  A method for validating an association between an invoice and an entity according to another embodiment of the present invention includes training a classifier based on the type of invoice associated with the first entity, the first entity and the other Accessing a plurality of invoices labeled to be associated with at least one of the entities, performing a document classification technique on the invoice using a classifier, and associated with the first entity Outputting at least one identifier of invoices having a high probability of not doing so.

  A method for managing medical records according to another embodiment of the present invention includes training a classifier based on a medical diagnosis, accessing a plurality of medical records, and using the classifier to classify documents with respect to the medical records. And outputting an identifier of at least one of the medical records that have a low probability of being associated with a medical diagnosis.

  A face recognition method according to another embodiment of the present invention includes receiving a seed image of at least one labeled face having a known confidence level, receiving an unlabeled image, and at least one predetermined cost factor. A transductive classifier by an iterative calculation using at least one predetermined cost factor, at least one seed image, and an unlabeled image, with the cost factor adjusted as a function of the expected label value for each And, after at least some iteration, storing a confidence score for the unlabeled seed image, and identifying an identifier for the unlabeled image with the highest confidence score for the user, another system, and another process. And outputting to at least one of them.

  A method for analyzing a prior art document according to another embodiment of the present invention includes training a classifier based on a search query, accessing a plurality of prior art documents, and using the classifier to prior art document. Performing a document classification technique for at least some of the documents and outputting identifiers for at least some of the documents based on the classification of the prior art documents.

  A method for adapting a patent classification to shifting document content according to another embodiment of the invention includes receiving at least one labeled seed document, receiving an unlabeled document, and at least one seed document and label. Using a non-document to train a transductive classifier, using the classifier to classify unlabeled documents having a confidence level above a predetermined threshold into a plurality of existing categories, and using a classifier Classifying unlabeled documents having a confidence level below a predetermined threshold into at least one new category, and using a classifier to categorize at least some of the categorized documents with the existing category and at least one Re-categorize into two new categories and categorized document identifiers , Including the steps of outputting at least one of another system, and another process.

  According to another embodiment of the present invention, a method for matching a document to a claim includes training a classifier and accessing a plurality of documents based on at least one claim of a patent document or patent application document. And using a classifier to perform a document classification technique for at least a portion of the document, and outputting an identifier for at least a portion of the document based on the classification of the document.

  A method for classifying a patent document or patent application document according to another embodiment of the present invention comprises the steps of training a classifier based on a plurality of documents known to exist in a particular patent classification; Receiving at least a portion of an application document, performing a document classification technique on at least a portion of a patent document or patent application document using a classifier, and outputting a classification of the patent document or patent application document The document classification method is a yes / no classification method.

  A method for classifying a patent document or patent application document according to another embodiment of the invention uses a classifier trained based on at least one document associated with a particular patent classification, and at least a patent document or patent application document. For some, including performing a document classification technique that is a yes / no classification technique and outputting a classification of a patent document or patent application document.

  A method for adapting to document content shifting according to another embodiment of the present invention includes receiving at least one labeled seed document, receiving an unlabeled document, and receiving at least one predetermined cost factor. Training a transductive classifier using at least one predetermined cost factor, at least one seed document, and an unlabeled document; and using the classifier, no label having a confidence level above a predetermined threshold Classifying a document into a plurality of categories; reclassifying at least a portion of the categorized document into categories using a classifier; and identifying an identifier of the categorized document to a user, another system, And outputting to at least one of the other processes.

  A method for separating documents according to another embodiment of the present invention includes receiving labeled data, receiving a series of unlabeled documents, and transduction based on labeled data and unlabeled documents, Adapting the probabilistic classification rules, updating the weights used for document classification according to the probabilistic classification rules, determining the separation positions in the series of documents, and indicating the separation positions in the determined series. Outputting to at least one of the user, another system, and another process, and setting a flag in the document that correlates with the sign.

  A document search method according to another embodiment of the present invention includes a search query for at least one of: receiving a search query; retrieving a document based on the search query; outputting a document; and Receiving a user input label indicating the relevance of the document to the document, training a classifier based on the search query and the user input label, and using the classifier to reclassify the document, Performing a document classification technique and outputting at least some identifiers of the document based on the document classification.
  The present invention also provides the following items, for example.
(Item 1)
A method of data classification in a computer-based system, comprising:
Receiving labeled data points, wherein each of the labeled data points is a training example for a data point for which the data point is to be included in a specified category or excluded from a specified category. Having at least one label indicating whether the training is for an example data point;
Receiving unlabeled data points;
Receiving at least one predetermined cost factor for the labeled and unlabeled data points;
Training a transductive classifier using maximum entropy discrimination (MED) by iterative computation using the at least one cost factor and the labeled and unlabeled data points as training examples, comprising: For each iteration, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point label is adjusted according to the estimate of the class membership probability of the data point;
Applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and an input data point;
Outputting the classification of the classified data points, or a derivative thereof, to at least one of a user, another system, and another process;
Including the method.
(Item 2)
Item 2. The method of item 1, wherein the function is an absolute value of the expected label of a data point.
(Item 3)
The method of item 1, further comprising receiving prior probability information for labeled and unlabeled data points.
(Item 4)
4. The method of item 3, wherein the transductive classifier learns using prior probability information of the labeled and unlabeled data.
(Item 5)
Using the labeled data and the unlabeled data as learning examples according to their expected labels, using a Gaussian prior distribution for the decision function parameters given the included training examples and excluded training examples The method of item 1, further comprising the further step of determining a decision function having a minimum KL divergence.
(Item 6)
The method of item 1, comprising the further step of determining a decision function having a minimum KL divergence using a polynomial prior for the decision function parameters.
(Item 7)
The method of item 1, wherein the iterative step of training a transductive classifier is repeated until convergence of data values is reached.
(Item 8)
8. The method of item 7, wherein convergence is reached when a change in the decision function of the transductive classifier falls below a predetermined threshold.
(Item 9)
8. The method of item 7, wherein convergence is reached when the determined change in expected label value falls below a predetermined threshold.
(Item 10)
The method of item 1, wherein the labels of the included training examples have a value of +1 and the excluded training examples labels have a value of -1.
(Item 11)
2. The method of item 1, wherein the labels of the included examples are mapped to a first number and the labels of the excluded examples are mapped to a second number.
(Item 12)
Storing the labeled data points in a memory of a computer;
Storing the unlabeled data points in a memory of a computer;
Storing the input data points in a memory of a computer;
Storing at least one predetermined cost factor of the labeled and unlabeled data points in a memory of a computer;
The method according to item 1, further comprising:
(Item 13)
A method of data classification comprising providing computer-executable program code that is deployed and executed on a computer system, comprising:
The program code is
At least one indicating whether each labeled data point is a training example for a data point to be included in the specified category or a training point for a data point excluded from the specified category Accessing the labeled data point stored in the memory of the computer having one label;
Access unlabeled data points from computer memory,
Accessing at least one predetermined cost factor of the labeled and unlabeled data points from the memory of the computer;
Using the at least one stored cost factor and stored labeled data points and stored unlabeled data points to train a maximum entropy identification (MED) transductive classifier by iterative calculations, For iterations, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, the prior probability of the data point label is adjusted according to the estimate of the class membership probability of the data point,
Applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and the input data point;
Outputting the classification of the classified data points, or a derivative thereof, to at least one of a user, another system, and another process;
With instructions for
Method.
(Item 14)
14. The method of item 13, wherein the function is the absolute value of the expected label of a data point.
(Item 15)
14. The method of item 13, further comprising accessing prior probability information of labeled and unlabeled data points stored in a computer memory.
(Item 16)
16. The method of item 15, wherein for each iteration, the prior probability information is adjusted according to an estimate of a data point class membership probability.
(Item 17)
Utilizing the labeled and unlabeled data as learning examples according to their expected labels and having a minimum KL divergence for the pre-distribution of decision function parameters given the included training examples and excluded training examples 14. The method of item 13, further comprising instructions for determining a decision function.
(Item 18)
14. The method of item 13, wherein the iterative step of training a transductive classifier is repeated until convergence of data values is reached.
(Item 19)
Item 19. The method of item 18, wherein convergence is reached when the change in the decision function of the transductive classification falls below a predetermined threshold.
(Item 20)
19. The method of item 18, wherein convergence is reached when the determined change in expected label value falls below a predetermined threshold.
(Item 21)
14. The method of item 13, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1.
(Item 22)
14. The method of item 13, wherein the labels of the included examples are mapped to a first number and the labels of the excluded examples are mapped to a second number.
(Item 23)
A data processing device, the device comprising:
(I) Whether each labeled data point is a training example for a data point that should be included in the specified category or a training example for a data point excluded from the specified category Storing at least one labeled data point, (ii) an unlabeled data point, and (iii) at least one predetermined cost factor for the labeled and unlabeled data point, at least One memory,
Using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples, a transductive maximum entropy identification (MED) is used for the transductive classifier. A transductive classifier training device for iterative teaching, wherein at each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the label of the data point is A training device that is adjusted according to an estimate of the class membership probability of the data points;
With
A classifier trained by the transductive classifier training device is used to classify at least one of the unlabeled data points, the labeled data points, and the input data points;
The classification of the classified data points, or a derivative thereof, is output to at least one of a user, another system, and another process;
apparatus.
(Item 24)
24. The apparatus of item 23, wherein the function is an absolute value of the expected label of a data point.
(Item 25)
24. The apparatus of item 23, wherein the memory also stores prior probability information for labeled and unlabeled data points.
(Item 26)
26. The apparatus of item 25, wherein in each iteration of the MED calculation, the prior probability information is adjusted according to an estimate of a data point class membership probability.
(Item 27)
Using the labeled and unlabeled data as learning examples according to their expected labels, the included training examples and excluded training examples have a minimum KL divergence for given decision function prior distributions 24. The apparatus of item 23, further comprising a processor for determining a decision function.
(Item 28)
Item 24. The apparatus according to Item 23, further comprising means for determining convergence of the data value and terminating the calculation simultaneously with the determination of convergence.
(Item 29)
29. The apparatus of item 28, wherein convergence is reached when a change in a decision function of the transductive classifier calculation falls below a predetermined threshold.
(Item 30)
29. The apparatus of item 28, wherein convergence is reached when the determined change in expected label value falls below a predetermined threshold.
(Item 31)
24. The apparatus of item 23, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1.
(Item 32)
24. The apparatus of item 23, wherein the labels of the included examples are mapped to a first number and the labels of the excluded examples are mapped to a second number.
(Item 33)
A product comprising a computer readable program storage medium, which unambiguously embodies one or more programs of instructions executable by a computer to perform a data classification method, the method comprising:
Each labeled data point indicates whether it is a training example for a data point to be included in a specified category or a training example for a data point excluded from a specified category, at least Receiving the labeled data point having one label;
Receiving unlabeled data points;
Receiving at least one predetermined cost factor of the labeled and unlabeled data points;
Training a transductive classifier by iterative maximum entropy identification (MED) calculation using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples. Where in each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point is adjusted according to the estimate of the class membership probability of the data point. Steps,
Applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and an input data point;
Outputting the classification of the classified data points, or a derivative thereof, to at least one of a user, another system, and another process;
Including
Product.
(Item 34)
34. A product according to item 33, wherein the function is an absolute value of the expected label of a data point.
(Item 35)
34. The product of item 33, wherein the method further comprises storing prior probability information for labeled and unlabeled data points in a computer memory.
(Item 36)
36. The product of item 35, wherein in each iteration of the MED calculation, the prior probability information is adjusted according to an estimate of a data point class membership probability.
(Item 37)
The method uses the labeled and unlabeled data as learning examples according to their expected labels, and uses a minimum for a pre-distribution of decision function parameters given the included training examples and excluded training examples. 34. The product of item 33, comprising the further step of determining said decision function having KL divergence.
(Item 38)
The product of item 33, wherein the iterative step of training the transductive classifier is repeated until convergence of the data values is reached.
(Item 39)
40. The product of item 38, wherein convergence is reached when a change in the decision function of the transductive classification falls below a predetermined threshold.
(Item 40)
39. The product of item 38, wherein convergence is reached when the determined change in expected label value falls below a predetermined threshold.
(Item 41)
34. The product of item 33, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1.
(Item 42)
34. The product of item 33, wherein the label of the included example is mapped to a first number and the label of the excluded example is mapped to a second number.
(Item 43)
A method for classifying unlabeled data in a computer-based system, comprising:
Receiving a labeled data point, whether the data point is a training example for a data point to be included in a specified category or a training example for a data point excluded from a specified category Each of the labeled data points has at least one label indicating:
Receiving labeled and unlabeled data points;
Receiving pre-label probability information for labeled and unlabeled data points;
Receiving at least one predetermined cost factor for the labeled and unlabeled data points;
Determining an expected label for each labeled and unlabeled data point according to the prior probability of the label of the data point;
Until the data values substantially converge, the following substeps are taken:
Generating a scaled cost value for each unlabeled data point in proportion to the absolute value of the expected label of the data point;
Use the labeled and unlabeled data as training examples according to their expected labels to minimize KL divergence for the prior probability distribution of the decision function parameters given the included training examples and excluded training examples Training the classifier by calculating the decision function to:
Using the trained classifier to determine a classification score for the labeled and unlabeled data points;
Calibrating the output of the trained classifier to class membership probabilities;
Updating the prior probability of the label of the unlabeled data point according to the determined class membership probability;
Determining the probability distribution of the label and margin using maximum entropy identification (MED) using the updated prior probability of the label and the previously determined classification score;
Using the previously determined probability distribution of labels to calculate a new expected label;
Updating the expected label for each data point by interpolating the new expected label with the expected label from the previous iteration;
Repeating steps,
Outputting the classification of the input data points, or derivatives thereof, to at least one of a user, another system, and another process;
Including the method.
(Item 44)
44. A method according to item 43, wherein convergence is reached when a change in the decision function falls below a predetermined threshold.
(Item 45)
44. A method according to item 43, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold.
(Item 46)
45. The method of item 43, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1.
(Item 47)
Receiving at least one labeled seed document having a known confidence level for label assignment;
Receiving an unlabeled document;
Receiving at least one predetermined cost factor;
Training a transductive classifier by iteration using the at least one predetermined cost factor, the at least one seed document, and the unlabeled document, for each iteration of the computation, A step where the cost factor is adjusted as a function of the expected label value; and
Storing a confidence score for the unlabeled document after at least some of the iterations;
Outputting an identifier of the unlabeled document having the highest confidence score to at least one of a user, another system, and another process;
A method for classifying documents, including:
(Item 48)
48. The method of item 47, wherein the at least one seed document has a list of keywords.
(Item 49)
48. The method of item 47, wherein a confidence score is stored after each iteration and an identifier of the unlabeled document having the highest confidence score is output after each iteration.
(Item 50)
49. The method of item 47, further comprising receiving prior probabilities of data point labels for the labeled and unlabeled documents, wherein for each iteration of the calculation, the label of the data point. The method wherein the prior probabilities are adjusted according to an estimate of the class membership probability of the data points.
(Item 51)
Receiving documents related to legal matters;
Performing a document classification technique on the document;
Outputting an identifier of at least a portion of the document based on the classification of the document;
Analyzing documents related to legal disclosure procedures, including:
(Item 52)
52. A method according to item 51, wherein the document classification method includes transductive processing.
(Item 53)
Training a transductive classifier by iterative computation using at least one predetermined cost factor, at least one seed document, and a document associated with a legal matter, for each iteration of the computation, 53. The method of item 52, further comprising the step of adjusting the cost factor as a function of an expected label value.
(Item 54)
54. The method of item 53, further comprising receiving prior probabilities of data point labels for the labeled and unlabeled documents, for each iteration of the calculation, prior to the data point label priorities. The method, wherein the probability is adjusted according to an estimate of the class membership probability of the data point.
(Item 55)
52. A method according to item 51, wherein the document classification method includes support vector machine processing.
(Item 56)
52. A method according to item 51, wherein the document classification method includes a maximum entropy identification process.
(Item 57)
52. The method of item 51, further comprising outputting a representation representing a link between the documents.
(Item 58)
Receiving a plurality of labeled data items;
Selecting a subset of the data items for each of a plurality of categories;
Setting the uncertainty for the data items in each subset to approximately zero;
Setting the uncertainty for the data item not present in the subset to a predetermined value that is not substantially zero;
Training a transductive classifier by iterative computation using the uncertainty, data items in the subset, and the data items not in the subset as training examples;
Applying the trained classifier to each of the labeled data items to classify each of the data items;
Outputting the classification of the input data item, or a derivative thereof, to at least one of a user, another system, and another process;
A way to organize data, including
(Item 59)
59. The method of item 58, wherein the subset is selected randomly.
(Item 60)
59. The method of item 58, wherein the subset is selected and verified by a user.
(Item 61)
59. The method of item 58, further comprising changing the label of at least some of the data items based on the classification.
(Item 62)
59. The method of item 58, wherein after classification of the data item, an identifier of the data item having a confidence level below a predetermined threshold is output to the user.
(Item 63)
Training a classifier based on the type of invoice associated with the first entity;
Accessing a plurality of invoices labeled to be associated with at least one of the first and other entities;
Using the classifier to perform a document classification technique on the invoice;
Outputting an identifier of at least one of the invoices having a high probability not associated with the first entity;
A method for verifying the relationship between an invoice and an entity, including
(Item 64)
64. A method according to item 63, wherein the document classification method includes transductive processing.
(Item 65)
65. The method of item 64, wherein the classifier is a transductive classifier, the method using an iterative calculation using at least one predetermined cost factor, at least one seed document, and the invoice. Training the transductive classifier, wherein for each iteration of the calculation, the cost factor is adjusted as a function of an expected label value, and using the trained classifier Classifying the invoice.
(Item 66)
68. The method of item 65, further comprising receiving a prior probability of a data point label for the seed document and invoice, wherein for each iteration of the calculation, the prior probability of the data point label. Is adjusted according to an estimate of the class membership probability of the data points.
(Item 67)
64. A method according to item 63, wherein the document classification method includes support vector machine processing.
(Item 68)
64. A method according to item 63, wherein the document classification method includes a maximum entropy identification process.
(Item 69)
Training a classifier based on a medical diagnosis;
Accessing multiple medical records;
Using the classifier to perform a document classification technique on the medical record;
Outputting an identifier of at least one of the medical records having a low probability associated with the medical diagnosis;
A method of managing medical records.
(Item 70)
70. The method of item 69, wherein the document classification technique includes transductive processing.
(Item 71)
71. The method of item 70, wherein the classifier is a transductive classifier, wherein the transductive classifier is repetitively calculated using at least one predetermined cost factor, at least one seed document, and the medical record. Training a classifier, wherein for each iteration of the calculation, the cost factor is adjusted as a function of an expected label value, and the trained classifier to classify the medical record Using the method.
(Item 72)
72. The method of item 71, further comprising receiving a prior probability of a data point label for the seed document and medical record, wherein for each iteration of the calculation, the prior probability of the data point label. Is adjusted according to an estimate of the class membership probability of the data points.
(Item 73)
70. A method according to item 69, wherein the document classification technique includes support vector machine processing.
(Item 74)
70. The method of item 69, wherein the document classification technique includes a maximum entropy identification process.
(Item 75)
Receiving at least one face labeled seed image having a known confidence level;
Receiving an unlabeled image;
Receiving at least one predetermined cost factor;
Training a transductive classifier by iteration using the at least one predetermined cost factor, the at least one seed image, and the unlabeled image, for each iteration of the computation, A step where the cost factor is adjusted as a function of the expected label value; and
Storing a confidence score for an unlabeled seed image after at least some of the iterations;
Outputting an identifier of the unlabeled image having the highest confidence score to at least one of a user, another system, and another process;
A face recognition method.
(Item 76)
76. The method of item 75, wherein the at least one seed image has a label that indicates whether the image is included in a specified category.
(Item 77)
76. The method of item 75, wherein a confidence score is stored after each of the iterations and an identifier of the unlabeled image having the highest confidence score is output after each iteration.
(Item 78)
76. The method of item 75, further comprising receiving prior probabilities of data point labels for the labeled and unlabeled images, wherein for each iteration of the calculation, the label of the data point. The method wherein the prior probabilities are adjusted according to an estimate of the class membership probability of the data points.
(Item 79)
Receiving a third unlabeled image of a face; comparing the third unlabeled image with at least a portion of the image having the highest confidence score; and a face of the third unlabeled image 76. The method of item 75, further comprising: outputting an identifier of the third unlabeled image if the confidence of is the same as the face of the seed image.
(Item 80)
Training a classifier based on a search query;
Accessing a plurality of prior art documents;
Using the classifier to perform a document classification technique on the prior art document;
Outputting an identifier of at least a portion of the prior art document based on the classification of the prior art document;
A method for analyzing a prior art document including:
(Item 81)
81. A method according to item 80, wherein the document classification technique includes transductive processing.
(Item 82)
82. The method of item 81, wherein the classifier is a transductive classifier, wherein the transductor is repetitively calculated using at least one predetermined cost factor, at least one seed document, and the prior art document. Training an active classifier, wherein for each iteration of the calculation, the cost factor is adjusted as a function of an expected label value, and the trained to classify the prior art document Using a classifier.
(Item 83)
83. The method of item 82, further comprising receiving prior probabilities of data point labels for the seed document and the prior art document, wherein for each iteration of the calculation, prior probabilities of data point labels. Is adjusted according to an estimate of the class membership probability of the data points.
(Item 84)
81. The method of item 80, wherein the search query includes at least a portion of patent disclosure information.
(Item 85)
81. The method of item 80, wherein the search query includes at least a portion of an item retrieved from a patent document or patent application document.
(Item 86)
81. The method of item 80, wherein the search query includes at least a portion of a patent document or a summary of patent application documents.
(Item 87)
81. The method of item 80, wherein the search query includes at least a portion of a summary retrieved from a patent document or patent application document.
(Item 88)
81. A method according to item 80, wherein the document classification method includes support vector machine processing.
(Item 89)
81. The method of item 80, wherein the document classification technique includes a maximum entropy identification process.
(Item 90)
81. A method according to item 80, wherein the prior art document is a patent office public document.
(Item 91)
The method of item 80, further comprising outputting a representation of a link between the documents.
(Item 92)
81. The method of item 80, further comprising outputting a relevance score for at least a portion of the prior art document based on the prior art document classification.
(Item 93)
Receiving at least one labeled seed document;
Receiving an unlabeled document;
Training a transductive classifier using the at least one seed document and the unlabeled document;
Using the classifier to classify the unlabeled documents having a confidence level above a predetermined threshold into a plurality of existing categories;
Using the classifier to classify the unlabeled document having a confidence level below a predetermined threshold into at least one new category;
Reclassifying at least a portion of the categorized document into the existing category and the at least one new category using the classifier;
Outputting an identifier of the categorized document to at least one of a user, another system, and another process;
To adapt patent classification to document content shifts.
(Item 94)
94. The method of item 93, wherein the classifier is a transductive classifier, and trains the transductive classifier by iterative calculation using at least one predetermined cost factor, a search query, and the document. And, for each iteration of the calculation, the cost factor is adjusted as a function of an expected label value; and using the trained classifier to classify the document; Further comprising a method.
(Item 95)
95. The method of item 94, further comprising receiving a prior probability of a data point label for the search query and document, wherein for each iteration of the calculation, the prior probability of the data point label is A method that is adjusted according to an estimate of the class membership probability of the data points.
(Item 96)
94. A method according to item 93, wherein the document classification method includes support vector machine processing.
(Item 97)
94. A method according to item 93, wherein the document classification method includes a maximum entropy identification process.
(Item 98)
94. A method according to item 93, wherein the unlabeled document is a patent application document.
(Item 99)
94. The method of item 93, wherein the at least one seed document is selected from the group consisting of patent documents and patent application documents.
(Item 100)
Training a classifier based on at least one item of a patent document or patent application document;
Accessing multiple documents;
Using the classifier to perform a document classification technique on at least a portion of the document;
Outputting an identifier of at least a portion of the document based on the classification of the document;
A method of matching documents to items, including
(Item 101)
101. The method of item 100, further comprising outputting a relevance score for at least a portion of the document based on the classification of the document.
(Item 102)
101. A method according to item 100, wherein the document is a prior art document.
(Item 103)
101. A method according to item 100, wherein the document describes a product.
(Item 104)
Training a classifier based on a plurality of documents known to exist in a particular patent classification;
Receiving at least a portion of a patent document or patent application document;
Performing a document classification technique on the at least part of the patent document or patent application document using the classifier;
Outputting the classification of the patent document or patent application document;
A method for classifying patent documents or patent application documents, comprising:
The document classification technique is a yes / no classification technique.
(Item 105)
105. The method of item 104, wherein the document is selected from the group consisting of patent documents and patent application documents.
(Item 106)
106. The method of item 105, wherein the at least part of the patent document or patent application document includes at least a part of an item retrieved from the patent document or patent application document.
(Item 107)
106. The method of item 105, wherein the at least part of the patent document or patent application document includes at least a part of a summary of the patent document or patent application document.
(Item 108)
106. The method of item 105, wherein the at least part of the patent document or patent application document includes at least a part of a summary taken from the patent document or patent application document.
(Item 109)
Performing a document classification technique on at least a portion of a patent document or patent application document using a classifier trained based on at least one document associated with a particular patent classification, the document classification technique comprising: Yes / No formula classification method, step,
Outputting the classification of the patent document or patent application document;
Classifying patent documents or patent application documents, including:
(Item 110)
110. The method of item 109, further comprising repeating the method with different classifiers trained based on a plurality of documents known to exist in the second patent classification.
(Item 111)
110. The method of item 109, wherein the at least part of the patent document or patent application document includes at least a part of an item retrieved from the patent document or patent application document.
(Item 112)
110. The method of item 109, wherein the at least part of the patent document or patent application document includes at least a part of a summary of the patent document or patent application document.
(Item 113)
110. The method of item 109, wherein the at least part of the patent document or patent application document includes at least a part of a summary taken from the patent document or patent application document.
(Item 114)
Receiving at least one labeled seed document;
Receiving an unlabeled document;
Receiving at least one predetermined cost factor;
Training a transductive classifier using the at least one predetermined cost factor, the at least one seed document, and the unlabeled document;
Using the classifier to classify the unlabeled documents having a confidence level above a predetermined threshold into a plurality of categories;
Outputting an identifier of the categorized document to at least one of a user, another system, and another process;
To adapt to shifts in document content.
(Item 115)
115. The method of item 114, further comprising transferring unlabeled documents having a confidence level below the predetermined threshold to one or more new categories.
(Item 116)
Training the transductive classifier by iterative computation using at least one predetermined cost factor, the at least one seed document, and the unlabeled document, for each iteration of the computation, 115. The method of item 114, further comprising: a cost factor is adjusted as a function of expected label value; and using the trained classifier to classify the unlabeled document.
(Item 117)
118. The method of item 116, further comprising receiving prior probabilities of data point labels for the seed document and unlabeled document, wherein for each iteration of the calculation, the label of the data point. The method wherein the prior probabilities are adjusted according to an estimate of the class membership probability of the data points.
(Item 118)
119. The method of item 114, wherein the unlabeled document is a customer complaint, further comprising linking a product change with a customer complaint.
(Item 119)
115. The method of item 114, wherein the unlabeled document is an invoice.
(Item 120)
Receiving labeled data; and
Receiving a sequence of unlabeled documents;
Adapting probabilistic classification rules using transduction based on the labeled data and the unlabeled document;
Updating weights used for document separation according to the probabilistic classification rules;
Determining separation positions in the sequence of documents;
Outputting the indicator of the determined separation location in the series to at least one of a user, another system, and another process;
Setting a flag in the document that correlates with the indicator;
A method for separating documents, including:
(Item 121)
Receiving a search query;
Retrieving a document based on the search query;
Outputting the document;
Receiving a user input label for at least a portion of the document, the label indicating relevance of the document to the search query;
Training a classifier based on the search query and the user input label;
Performing a document classification technique on the document using the classifier to reclassify the document;
Outputting an identifier of at least a portion of the document based on the classification of the document;
Document retrieval method including
(Item 122)
124. A method according to item 121, wherein the document classification technique includes transductive processing.
(Item 123)
123. The method of item 122, wherein the classifier is a transductive classifier, wherein the transductive classifier is repetitively calculated using at least one predetermined cost factor, the search query, and the document. Training, wherein for each iteration of the calculation, the cost factor is adjusted as a function of an expected label value; and using the trained classifier to classify the document; Further comprising.
(Item 124)
124. The method of item 123, further comprising: receiving a prior probability of a data point label for the search query and document, wherein for each iteration of the calculation, the prior probability of the label of the data point is A method that is adjusted according to an estimate of the class membership probability of the data points.
(Item 125)
124. A method according to item 121, wherein the document classification method includes support vector machine processing.
(Item 126)
124. A method according to item 121, wherein the document classification method includes a maximum entropy identification process.
(Item 127)
122. The method of item 121, wherein the reclassified document is output and the document with the highest confidence is output first.

  The following description is the best mode presently contemplated for carrying out the invention. This description is made for the purpose of illustrating the general principles of the invention and is not intended to limit the inventive concepts claimed herein. Furthermore, the particular features described herein can be used in combination with the other features described in each of the various possible combinations and permutations.

  Unless defined otherwise specifically in the specification, all terms are intended to include the meanings provided by the specification and the meanings understood by those of ordinary skill in the art and defined in dictionaries, technical books, etc. Is given the widest interpretation possible.

(Character classification)
Interest and need for character data classification is particularly strong, and several classification techniques have been employed. Below, the classification method of character data is examined.

  In order to improve the usefulness and intelligence of classification methods, machines such as computers are needed to classify (or recognize) objects into ever-increasing content. For example, a computer can classify handwritten or scanned numbers and letters using optical character recognition, and can classify images such as faces, fingerprints, fighters, etc. using pattern recognition. Alternatively, sound, voice, etc. can be classified using speech recognition.

  Machines have also been required to classify character information objects such as computer files or documents consisting of characters. There are various and important applications for character classification. For example, character classification can be used to organize character information objects, for example, into a hierarchical structure of a given class or category. This approach simplifies the discovery (or navigation to) character information objects associated with a particular subject. Character classification can be used to send character information objects to the appropriate people or places. In this way, the information service industry sends textual information objects covering a wide variety of subjects (eg business, sports, stock market, football, specific companies, specific football teams) to people with various interests. Can do. Character classification can be used to filter character information objects so that individuals are not bothered by unwanted character content (such as junk mail or unwanted unsolicited mail, also called "spam"). . As can be seen from these few examples, there are many attractive and important uses for character classification.

(Rule-based classification)
In some cases, the character content needs to be classified with absolute certainty based on certain approved logic. A rule-based system can be used to perform such types of classification. Basically, rule-based systems use production rules of the form:
If (condition), it is (facts).
Here, the condition may include whether the character information includes a specific word or phrase, whether it has a specific syntax, or whether it has a specific attribute. For example, if the character content has the word “finished”, the phrase “Nasdaq” and a number, it is classified as a character for “stock market”.

  During the last decade or so, other types of classifiers have been increasingly used. Although these classifiers do not use static and predefined logic like rule-based classifiers, they have shown performance in many applications over rule-based classifiers. Such a classifier typically includes a learning element and an execution element. Such classifiers can include neural networks, Bayesian networks, and support vector machines. Each of these classifiers is known, but for the convenience of the reader, each is briefly introduced below.

(Classifier with learning and execution elements)
As just mentioned at the end of the previous section, classifiers with learning and execution elements outperform rule-based classifiers in many applications. To reiterate, these classifiers can include neural networks, Bayesian networks, and support vector machines.

(neural network)
A neural network is basically a hierarchical arrangement over multiple layers of identical processing elements, also called neurons. Each neuron can have more than one input, but only one output. Each neuron input is weighted by a coefficient. The output of a neuron is usually a function of the sum of the weighted inputs and the bias value. This function, also called the activation function, is generally a sigmoid function. That is, the activation function is sigmoidal and can be monotonically increasing and asymptotically fixed values (eg, +1, 0, −1) as its input (s) approaches positive or negative infinity, respectively. ). The sigmoid function and the individual neural weights and bias values determine the neuron's response or “sensitivity” to the input signal.

  In a hierarchical arrangement of neurons, the output of neurons in one layer can be distributed as an input to one or more neurons in the next layer. A typical neural network may include an input layer and two separate layers: an input layer, an intermediate neuron layer, and an output neuron layer. Note that the nodes in the input layer are not neurons. Rather, an input layer node has only one input and basically feeds that input unprocessed to the next layer input. For example, if a neural network is used to recognize numbers in a 20 × 15 pixel array, the input layer may have 300 neurons (ie, one for each pixel in the input) and the output array You can have 10 neurons (ie, one for each of the 10 numbers).

  The use of a neural network as a whole includes two successive stages. Initially, the network is initialized and trained on known inputs with known output values (or classifications). Once the neural network is trained, it can then be used to classify unknown inputs. The neural network can be initialized by setting neuron weights and biases to random values, typically generated from a Gaussian distribution. The neural network is then trained using a series of inputs with known outputs (or classifications). As the training inputs are fed into the neural network, the neuron weights and bias values are adjusted so that the neural network output for each individual training pattern approaches or matches the known output (eg, According to the known backpropagation method). Basically, a gradient descend in the weight space is used to minimize the output error. In this way, learning with continuous training input converges towards a local optimal solution for weights and bias. That is, the weight and bias are adjusted to minimize the error.

  In practice, this system is usually not trained until it converges to an optimal solution. Otherwise, the system will be “overtrained” and as a result, the system will be overly specialized for training data, classifying inputs that are somewhere different from the inputs in the training set. It may become unsuitable. Thus, at various points during the training period, the system is tested using a set of validation data. When the performance of this system using the verification set no longer improves, the training is stopped.

  Once training is complete, the neural network can be used to classify unknown inputs based on weights and biases calculated during training. If the neural network can reliably classify unknown inputs, one of the neuron outputs in the output layer will be much higher than the others.

(Bayesian network)
In general, Bayesian networks use hypotheses as an intermediate stage between data (eg, feature vector input) and prediction (eg, classification). Given the data, the probability of each hypothesis (“P (hypo | data)”) can be estimated. Using hypotheses posterior probabilities, predictions are made from hypotheses, and the individual predictions for each hypothesis are weighted. The probability of predicted X given data D is

で表され、ここで、Ｈ_ｉはｉ番目の仮説である。Ｄを所与とした場合のＨ_ｉの確率（Ｐ（Ｈ_ｉ｜Ｄ））を最大化する最も確からしい仮説は、最大事後仮説（または「Ｈ_ＭＡＰ」）と呼ばれ、

Where H _i is the i th hypothesis. The most probable hypothesis that maximizes the probability of H _i (P (H _i | D)) given D is called the maximum posterior hypothesis (or “H _MAP ”),

で表すことができる。
ベイズの定理を用いると、データＤを所与とした場合の仮説Ｈ．ｓｕｂ．ｉの確率は、

It can be expressed as
Using Bayes' theorem, the hypothesis H.D. sub. The probability of i is

で表すことができる。データＤの確率は固定されたままである。従って、Ｈ_ＭＡＰを求めるためには分子を最大化する必要がある。

It can be expressed as The probability of data D remains fixed. Therefore, it is necessary to maximize the molecule to determine H _MAP .

  The first term in the numerator represents the probability that the data would be observed with a hypothesis i. The second term represents the prior probability assigned to a given hypothesis i.

A Bayesian network includes variables and directed edges between variables, thereby defining a directed acyclic graph (or “DAG”). Each variable can take any finite number of mutually exclusive states. Parent variables B ₁ ,. . . For each variable A with B _n , a probability table (P (A | B ₁ ... B _n ) is attached.The structure of the Bayesian network is that each variable has a given parent variable. In this case, the assumption is that it is conditionally independent from the non-descendant of each variable.

  Assuming that the structure of the Bayesian network is known and the variables are observable, only a set of conditional probability tables need be learned. These tables can be estimated directly using statistics from a set of learning examples. If the structure is known but some variables are hidden, learning is similar to the neural network learning discussed above.

An example of a simple Bayesian network is introduced below. The variable “MML” may represent “my lawn moisture” and may have a “moist” state and a “dry” state. An MML variable may have a parent variable of “rain” and a parent variable of “my sprinkler is working”, each parent variable having a “yes” state and a “no” state. Another variable “MNL” may represent “my neighbor's lawn moisture” and may have a “wet” state and a “dry” state. MNL variables may share a parent variable of “rain”. In this example, the prediction can be whether my lawn is “moist” or “dry”. This prediction is hypothesis (i) “If it rains, my lawn will get wet with probability (x ₁ )”, hypothesis (ii) “If my sprinkler is working, my Lawn will get wet with probability (x ₂ ) ". The probability that it rained or that my sprinkler was working may depend on other variables. For example, if my neighbor's lawn is moist and the neighbor doesn't have a sprinkler, it's more likely that it's raining.

  As discussed above, conditional probability tables in Bayesian networks can be trained as in the case of neural networks. Useful, the learning process can be shortened by allowing provision of preliminary knowledge. Unfortunately, however, prior probabilities for conditional probabilities are usually unknown, in which case uniform prior probabilities are used.

  One embodiment of the present invention may perform at least one of two basic functions: generation of parameters for classifiers and classification of objects such as character information objects.

  Basically, parameters are generated for a classifier based on a set of training examples. From a set of training examples, a set of feature vectors can be generated. The features of a set of feature vectors can be reduced. The parameters to be generated may include predefined monotonic (eg sigmoid) functions and weight vectors. The weight vector can be determined by SVM training (or another known technique). Monotonic (eg, sigmoid) functions can be defined using optimization techniques.

  The character classifier may include a weight vector and a predefined monotonic (eg, sigmoid) function. Basically, the output of the character classifier of the present invention is

で表すことができる。ここで、
Ｏ_ｃ＝カテゴリｃに関する分類出力、
ｗ_ｃ＝カテゴリｃと関連付けられた重みベクトルのパラメータ、
ｘ＝未知の文字情報オブジェクトに基づく（縮約）特徴ベクトル、であり、
ＡおよびＢは、単調（例えばシグモイド）関数の調整パラメータである。

It can be expressed as here,
O _c = classification output for category c,
w _c = parameter of the weight vector associated with category c,
x = feature vector based on an unknown character information object,
A and B are adjustment parameters of a monotone (eg, sigmoid) function.

  The output calculation from equation (2) is faster than the output calculation from equation (1).

  Depending on the shape of the object to be classified, the classifier may: (i) convert the character information object into a feature vector, and (ii) reduce the feature vector to a feature vector having fewer elements. it can.

(Transductive machine learning)
Current state-of-the-art approaches in commercial automatic classification systems are either rule-based or use inductive machine learning, ie machine learning using manually labeled training examples. Both approaches generally require more manual setup effort compared to transductive methods. Solutions provided by rule-based systems or inductive approaches are static solutions that cannot adapt to drifting classification concepts without manual effort.

  Inductive machine learning formulates laws to attribute features or relationships to types based on tokens (ie, one or a few observations or experiences), or based on limited observations of recurring patterns Used to do. Inductive machine learning includes inferences from observed training examples to generate general rules, which are then applied to test examples.

  In particular, the preferred embodiment uses a transductive machine learning technique. Transductive machine learning is a powerful technique that does not suffer from these disadvantages.

  The transductive machine approach can learn from a very small set of labeled training examples while automatically adapting to drifting classification concepts and automatically modifying the labeled training examples. These advantages make transductive machine learning an interesting and valuable approach for a wide variety of commercial applications.

  Transduction methods learn patterns in data. By learning not only from labeled data but also from unlabeled data, it extends the concept of inductive learning. This allows the transduction method to learn patterns that are not captured or only partially captured in labeled data. As a result, in contrast to rule-based systems or inductive learning based systems, transduction methods can adapt to dynamically changing environments. This capability allows transduction methods to be used for document discovery, data organization, and especially for dealing with drifting classification concepts.

  The following is a description of one embodiment of transductive classification using a support vector machine (SVM) classification and maximum entropy identification (MED) framework.

(Support vector machine)
A support vector machine (SVM) is one method employed for character classification, and such a method introduces constraints on possible solutions using the concept of regularization theory, thereby solving a number of solution problems. Address points and the resulting generalization issues. For example, a binary SVM classifier selects the hyperplane that maximizes the margin as a solution from all hyperplanes that adequately separate the training data. Maximum margin normalization under the constraint that the training data is properly classified addresses the learning of the aforementioned problem of choosing an appropriate trade-off between generalization and memory. The constraints on the training data store the data, while normalization ensures proper generalization. Inductive classification learns from training examples with known labels, ie the class membership of all training examples is known. Inductive classification learns from known labels, while transductive classification determines classification rules from labeled and unlabeled data. An example of transductive SVM classification is shown in Table 1.

  (Principle of transductive SVM classification)

表１は、サポートベクタマシンを用いたトランスダクティブ分類の原理を示している。解は、ラベルなしデータの全てのあり得るラベル割り当てに関して、最大マージンをもたらす超平面（ｈｙｐｅｒｐｌａｎｅ）によって与えられる。あり得るラベル割り当ては、ラベルなしデータの数において指数関数的に増加し、実際に当てはまる解に対しては、表１のアルゴリズムを近似的に使用する必要がある。そのような近似の例は、Ｔ．Ｊｏａｃｈｉｍｓによる「Ｔｒａｎｓｄｕｃｔｉｖｅ ｉｎｆｅｒｅｎｃｅ ｆｏｒ ｔｅｘｔ ｃｌａｓｓｉｆｉｃａｔｉｏｎ ｕｓｉｎｇ ｓｕｐｐｏｒｔ ｖｅｃｔｏｒ ｍａｃｈｉｎｅｓ」、 Ｔｅｃｈｎｉｃａｌ ｒｅｐｏｒｔ、 Ｕｎｉｖｅｒｓｉｔａｅｔ Ｄｏｒｔｍｕｎｄ、 ＬＡＳ ＶＩＩＩ、 １９９９年に記載され（Ｊｏａｃｈｉｍｓ）ている。

Table 1 shows the principle of transductive classification using a support vector machine. The solution is given by the hyperplane that yields the maximum margin for all possible label assignments of unlabeled data. The possible label assignments increase exponentially in the number of unlabeled data, and for the solutions that apply in practice, the algorithm of Table 1 should be used approximately. An example of such an approximation is T.W. Joachims, “Transductive influence for text classification using supporting vector machines”, Technical report, Universitet Dortmund, 1999, J.

  The uniform distribution across the label assignments in Table 1 shows that unlabeled data points have a 1/2 probability of being a positive example of class and a 1/2 probability of being a negative example, ie y = + 1 ( The two possible label assignment probabilities of positive example) and y = −1 (negative example) are equal, meaning that the resulting label is zero. A zero label expectation can be determined by a fixed class prior probability equal to 1/2 or by a class prior probability that is a random variable with a uniform prior distribution, ie, an unknown class prior probability. Thus, in applications with a known class of prior probabilities not equal to 1/2, the algorithm can be improved by incorporating this additional information. For example, instead of using the uniform distribution for the label assignments in Table 1, it may be chosen to prioritize some label assignments over others according to class prior probabilities. However, the trade-off between a smaller margin solution with a plausible label assignment and a solution with a higher margin but less likely label assignment is difficult. The scale of label allocation probability and margin are different.

(Maximum entropy identification)
Another classification method, Maximum Entropy Discrimination (MED) method (see, eg, T. Jebara “Machine Learning Discriminative and General”, Kluwer Academic Publishers) (Jebara) is an assignment function normalization term and a decision function normalization term. Since both are derived from the prior probability distribution on the solution, and therefore both are on the same probabilistic measure, the problems associated with SVM are not encountered. Thus, if the prior probability of the class, and hence the prior probability of the label, is known, transductive MED classification allows the incorporation of prior label knowledge in a reasonable manner, and thus more than transductive SVM classification. Are better.

  Inductive MED classification assumes a prior distribution over the parameters of the decision function, a prior distribution over the bias term, and a prior distribution over the margin. The inductive MED classification method selects the closest distribution to the prior distribution as the final distribution on these parameters to obtain an estimated decision function that classifies the data points appropriately.

Formally, for example, for a linear classifier, this problem is formulated as follows: The distribution p (Θ), the bias distribution p (b), and the data point classification margin p (γ) with respect to the hyperplane parameter are set to the minimum cullback for each prior distribution p ₀ to which the combined probability distribution is combined. Ask to have a Liver Divergence KL, ie

は、下の制約条件に従う。

Follows the constraints below.

ここで、

here,

は、分離超平面の重みベクトルとｔ番目のデータ点の特徴ベクトルとのドット積である。ラベル割り当てｙ_ｔは既知でありかつ固定されているので、２値のラベル割り当てに対する事前分布は必要ではない。従って、帰納的ＭＥＤ分類をトランスダクティブＭＥＤ分類に一般化する直接的手法は、２値のラベル割り当てを、あり得るラベル割り当てに対する事前分布によって制約されるパラメータとして処理することである。トランスダクティブＭＥＤの一例を表２に示す。

Is the dot product of the weight vector of the separation hyperplane and the feature vector of the tth data point. Since the label allocation y _t is known and fixed, prior distribution for the label allocation binary is not necessary. Thus, a direct approach to generalizing inductive MED classification to transductive MED classification is to treat binary label assignments as parameters constrained by a prior distribution for possible label assignments. An example of a transductive MED is shown in Table 2.

  (Transductive MED classification)

ラベル付きデータに対しては、ラベルの事前分布はδ関数であり、従って、＋１または−１となるようにラベルを効果的に固定する。ラベルなしデータに対しては、ラベルの事前確率ｐ_０（ｙ）は、すべてのラベルなしデータ点に、ｐ_０（ｙ）の確率を有するｙ＝＋１の正のラベルおよび１−ｐ_０（ｙ）の確率を有するｙ＝−１の負のラベルを割り当ると仮定される。情報を提供しないラベルの事前確率（ｐ_０（ｙ）＝１／２）を仮定することで、上に論じたトランスダクティブＳＶＭ分類に類似したトランスダクティブＭＥＤ分類が得られる。

For labeled data, the label prior distribution is a δ function, thus effectively fixing the label to be +1 or −1. For unlabeled data, the label prior probability p ₀ (y) is the positive label of y = + 1 with a probability of p ₀ (y) and 1−p ₀ (y ) Is assumed to be assigned a negative label of y = −1 with probability of. By assuming a prior probability (p ₀ (y) = 1/2) of labels that do not provide information, a transductive MED classification similar to the transductive SVM classification discussed above is obtained.

  As in the case of transductive SVM classification, a practical implementation of such a MED algorithm needs to approximate the search across all possible label assignments. T.A. Jaakcola, M .; Meila and T.M. The “Maximum encyclopedia” by Jebara, Technical Report AITR-1668, Massachusetts Institute of Technology, Artificial Intelligence Laboratory, (Jaakkola) method described in 1999 is an approximation formula for expectation maximization (EM) formulation. Similarly, we have chosen to break down the procedure into two stages. This formulation has two problems to be solved. The first is similar to the M stage of the EM algorithm, similar to maximizing margins while properly classifying all data points according to the current best guess for label assignment. In the second stage, similar to the E stage, a new value related to the class attribution of each example is estimated using the classification result determined in the M stage. We refer to this second stage as label induction. The overall description is shown in Table 2.

The specific implementation of the Jaakcola approach referred to herein has a Gaussian distribution with mean zero and unit variance for hyperplane parameters and a mean zero and variance σ _b ² for bias parameters. For unlabeled data discussed above, the Gaussian distribution is a margin prior probability in the form exp [−c (1-γ)] where γ is the margin of the data points, c is the cost factor, and p ₀ ( Assume the prior probability of the binary label of y). For the following discussion of the transductive classification algorithm, Jaakcola, referred to herein, a prior probability of 1/2 label is assumed for reasons of simplification and not to lose generality.

  The label induction step determines the probability distribution of the label given a fixed probability distribution for the hyperplane parameters. Using the margin and label prior probabilities introduced above, the following objective function for the label induction stage is obtained (see Table 2).

ここで、λ_ｔはｔ回目の訓練例のラグランジュ乗数、ｓ_ｔは先のＭ段階で決定されたその分類スコア、ｃはコスト要因である。訓練例に関する合計の中の最初の２つの項はマージンの事前分布から導出されるが、それに対して、３番目の項はラベルの事前分布によって与えられる。

Here, the lambda _t Lagrange multipliers t-th training examples, s _t is the classification score determined in the previous M phase, c is a cost factor. The first two terms in the total for the training example are derived from the marginal prior, whereas the third term is given by the label prior.

を最大化することによってラグランジュ乗数が決定され、その結果として、ラベルなしデータに関するラベルの確率分布が決定される。式３から分かるように、各データ点は独立して目的関数に寄与する。従って、各ラグランジュ乗数は、他のすべてのラグランジュ乗数に関係なく決定され得る。例えば、その分類スコアの高い絶対値｜ｓ_ｔ｜を有するラベルなしデータ点の寄与を最大化するためには、小さいラグランジュ乗数λ_ｔが必要であるが、それに対して、小さい値｜ｓ_ｔ｜を有するラベルなしデータ点は、大きいラグランジュ乗数と共に、

To determine the Lagrangian multiplier and, as a result, the probability distribution of the labels for unlabeled data. As can be seen from Equation 3, each data point independently contributes to the objective function. Thus, each Lagrangian multiplier can be determined regardless of all other Lagrangian multipliers. For example, to maximize the contribution of an unlabeled data point with a high absolute value | s _t | of its classification score, a small Lagrange multiplier λ _t is required, whereas a small value | s _t | An unlabeled data point with a large Lagrange multiplier,

に対する寄与を最大化する。その一方では、ラベルなしデータ点の分類スコアｓおよびそのラグランジュ乗数λの関数としてのラベルなしデータ点の期待ラベル＜ｙ＞は、

Maximize the contribution to. On the other hand, the expected label <y> of the unlabeled data point as a function of the classification score s of the unlabeled data point and its Lagrange multiplier λ is

となる。
図１に、ｃ＝５およびｃ＝１．５のコスト要因を用いた分類スコアｓの関数としての期待ラベル＜ｙ＞を示す。図１の生成に用いたラグランジュ乗数は、ｃ＝５およびｃ＝１．５のコスト要因を用いて式３を解くことによって決定された。図１から分かるように、マージンの外側、すなわち｜ｓ｜＞１のラベルなしデータ点は、ゼロに近い期待ラベル＜ｙ＞を有しており、マージンに近い、すなわち｜ｓ｜≒１のデータ点は、最も高い期待ラベル絶対値をもたらし、超平面に近い、すなわち｜ｓ｜＜∈のデータ点は、｜＜ｙ＞｜＜∈をもたらす。｜ｓ｜→∞に対して＜ｙ＞→０というこの非直感的ラベル割り当ての理由は、分類上の制約が満たされる限りはできるだけ事前分布の近傍にとどまろうとする、選択された識別的手法にある。これは、表２の既知の手法によって選択された近似式のアーチファクトではなく、すなわち、あり得る全ラベル割り当てを網羅的に検索し、従って、大域的最適解を求めることを保証するするアルゴリズムがまた、マージンの外側のラベルなしデータにもゼロに近いかまたはゼロに等しい期待ラベルを割り当てる。上に述べたように、ここでもまた、識別的観点からそれが期待される。マージンの外側のデータ点は、例を分離するのには重要ではなく、従って、これらのデータ点のすべての個々の確率分布は、それらの事前確率分布に戻る。

It becomes.
FIG. 1 shows an expected label <y> as a function of the classification score s using cost factors of c = 5 and c = 1.5. The Lagrangian multiplier used to generate FIG. 1 was determined by solving Equation 3 using a cost factor of c = 5 and c = 1.5. As can be seen from FIG. 1, the unlabeled data point outside the margin, ie, | s |> 1, has an expected label <y> close to zero and is close to the margin, ie, | s | ≈1. A point yields the highest expected label absolute value, and a data point that is close to the hyperplane, ie, | s | <ε yields | <y> | <ε. The reason for this non-intuitive label assignment, <y> → 0 for | s | → ∞, is that the selected discriminative approach tries to stay as close as possible to the prior distribution as long as the classification constraints are met. is there. This is not an artifact of the approximate expression selected by the known method of Table 2, ie, an algorithm that ensures an exhaustive search of all possible label assignments, and thus a global optimal solution is also found. Allocate unlabeled data outside the margin to expected labels close to or equal to zero. As mentioned above, this is also expected here from a discriminative point of view. Data points outside the margin are not important to isolate the example, so all individual probability distributions of these data points revert to their prior probability distribution.

  The M stage of the Jaakcola transductive classification algorithm referenced herein determines the probability distribution for hyperplane parameters, bias terms, and data point margins that are closest to each prior distribution under the following constraints: To do.

ここで、ｓ_ｔはｔ回目のデータ点分類スコア、〈ｙ_ｔ〉はその期待ラベル、〈γ_ｔ〉はその期待マージンである。ラベル付きデータに対しては、期待ラベルは固定されており、＜ｙ＞＝＋１または＜ｙ＞＝−１である。ラベルなしデータに関する期待ラベルは、区間（−１、＋１）の中にあり、ラベル帰納段階で推定される。式５によれば、分類スコアは期待ラベルによってスケーリングされるので、ラベルなしデータは、ラベル付きデータよりも厳しい分類制約を満たす必要がある。さらに、図１を参照し、分類スコアの関数としての期待ラベルの依存性を所与とすると、分離超平面に近いラベルなしデータは、最も厳しい分類制約を有する。なぜならば、それらのスコアおよびそれらの期待ラベルの絶対値｜〈ｙ_ｔ〉｜が小さいからである。上述の事前分布を所与としたＭ段階の全目的関数は、

Here, _st is the t-th data point classification score, <y _t > is the expected label, and <γ _t > is the expected margin. For labeled data, the expected label is fixed and <y> = + 1 or <y> =-1. Expected labels for unlabeled data are in the interval (-1, +1) and are estimated at the label induction stage. According to Equation 5, since the classification score is scaled by the expected label, the unlabeled data needs to satisfy stricter classification constraints than the labeled data. Further, referring to FIG. 1, given the expected label dependence as a function of the classification score, unlabeled data close to the separation hyperplane has the most stringent classification constraints. This is because the absolute value | <y _t > | of those scores and their expected labels is small. The overall objective function of M stages given the above prior distribution is

となる。
第１項はガウスの超平面パラメータ事前分布から導出され、第２項はマージン事前正規化項、最後の項は、平均ゼロと分散σ_ｂ ^２とを有するガウス事前分布から導出されるバイアスの事前正規化項である。バイアス項に対する事前分布は、クラスの事前確率に対する事前分布として解釈され得る。従って、バイアスの事前分布に対応する正規化項は、正から負までの例の重みを制約する。式６によれば、バイアス項の寄与は、超平面上での正の例の一括プルと負の例の一括プルとが等しくなる場合に最小化される。バイアスの事前分布によるラグランジュ乗数に対する一括制約は、データ点の期待ラベルによって重み付けされ、従って、ラベル付きデータに対するよりもラベルなしデータに対する方が制約が少ない。従って、ラベルなしデータは、最終解に対して、ラベル付きデータよりも強い影響を与える能力を有する。

It becomes.
The first term is derived from a Gaussian hyperplane parameter prior distribution, the second term is a margin prenormalization term, the last term is a bias prior derived from a Gaussian prior distribution with mean zero and variance σ _b ^2. It is a normalization term. The prior distribution for the bias term can be interpreted as a prior distribution for the class prior probabilities. Thus, the normalization term corresponding to the bias prior distribution constrains the weight of the example from positive to negative. According to Equation 6, the bias term contribution is minimized when the positive example collective pull and the negative example collective pull on the hyperplane are equal. The bulk constraint on the Lagrangian multiplier due to the bias prior distribution is weighted by the expected labels of the data points and is therefore less constrained for unlabeled data than for labeled data. Thus, unlabeled data has the ability to influence the final solution more strongly than labeled data.

In summary, at the M stage of the Jaakcola transductive classification algorithm referred to herein, unlabeled data must meet stricter classification constraints than labeled data, and the accumulation of unlabeled data for the solution Weights are less constrained than for labeled data. Furthermore, unlabeled data with expected labels close to zero located within the current M-stage margin has the most impact on the solution. The net effect obtained from formulating the E and M stages in this manner is shown in FIG. 2 by applying this algorithm to the data set. This data set consists of two labeled examples: a negative example (x) located at x-position-1 and a positive example (+) located at +1, and -1 and +1 along the x-axis. And 6 unlabeled examples (O) located between. A cross (x) indicates a negative example with a label, a plus sign (+) indicates a positive example with a label, and a circle (◯) indicates unlabeled data. The various plots show the separation hyperplanes determined at various iterations of the M stage. The final solution chosen by the Jaakcola transductive MED classifier referenced herein misclassifies the positive labeled training example. FIG. 2 shows several iterations of the M stage. In the first iteration of the M stage, no unlabeled data is considered and the separation hyperplane is located at x = 0. One unlabeled data point with a negative x value is closer to this separation hyperplane than any other unlabeled data. In the next label induction stage, this unlabeled data point will be assigned the smallest | <y> |, so in the next M stage it will be the hyperplane for the positive labeled example. Has the greatest force to push. The specific shape (see FIG. 1) of the expected label <y> as a function of the classification score determined by the selected cost factor, combined with the specific spacing of unlabeled data points, is The separation hyperplane produces a bridging effect that gradually approaches toward a positive labeled example. Intuitively, at the M stage, the unlabeled data point closest to the latest separated hyperplane most determines the final position of the plane, and the farther away data point is a kind of myopia, which is less important. Eventually, the bias hyperdistribution term constrains the bulk pull of unlabeled data to less than the bulk pull of labeled data, so that the separation hyperplane moves beyond the positive labeled example and the final solution, That is, the 15th iteration of FIG. 2 is obtained, which misclassifies the positive labeled example. A bias variance of σ _b ² = 1 and a cost factor of c = 10 were used in FIG. With σ _b ² = 1, any cost factor in the range 9.8 <c <13 results in a final hyperplane that misclassifies one positive labeled example. Cost factors outside the interval 9.8 <c <13 result in a separation hyperplane somewhere between the two labeled examples.

  This instability of this algorithm is not limited to the example shown in FIG. 2, but also during the application of the Jaakcola method referenced herein to the real world containing Reuters datasets known to those skilled in the art. Have been experienced. The inherent instability of the method described in Table 2 is a major drawback of this implementation, limiting its general applicability, but the Jaakcola method can be implemented in some embodiments of the present invention. .

  One preferred approach of the present invention employs transductive classification using a maximum entropy identification (MED) framework. While various embodiments of the present invention are applicable to classification, other MED learning problems using transduction, including but not limited to, transductive MED regression and graphical models. It should be understood that this is also applicable.

  Maximum entropy discrimination methods constrain and reduce possible solutions by assuming prior probability distributions for the parameters. The final solution is the expected value of all possible solutions according to the probability distribution closest to the assumed prior probability distribution, under the constraint that the expected solution adequately describes the training data. The prior probability distribution above the solution maps to the normalization term. That is, a specific normalization is selected by selecting a specific prior distribution.

  The discriminative estimation applied by the support vector machine is effective when learning from a few examples. This method and apparatus of one embodiment of the present invention has this as well as a support vector machine and does not attempt to estimate more parameters than necessary to solve a given problem, resulting in a sparse solution. Bring. This is in contrast to generative model estimation, which attempts to explain the underlying process and generally requires more statistical data than discriminative estimation. On the one hand, generative models are more versatile and can be applied to a wider variety of problems. Furthermore, generative model estimation can directly include conventional knowledge. The method and apparatus of an embodiment of the present invention using maximum entropy discrimination bridges the gap between purely discriminatory, eg, support vector machine learning and generative model estimation.

  The method of one embodiment of the present invention shown in Table 3 is an improved transductive MED classification algorithm that does not have the method instability issues discussed in Jaakcola referenced herein. The differences include, but are not limited to, in one embodiment of the present invention, that each data point has its own cost factor proportional to its label expected absolute value | <y> |. Furthermore, the prior probabilities of the labels of each data point are updated after each M stage according to the estimated class membership probability as a function of the distance of the data points to the decision function. The method of one embodiment of the present invention is illustrated in Table 3 below.

  (Improved transductive MED classification)

｜＜ｙ＞｜によってデータ点のコスト要因をスケーリングすることは、ラベルなしデータがラベル付きデータよりも超平面上でより強い累積プルを有し得るという問題を緩和する。なぜならば、ラベルなしデータのコスト要因は今やラベル付きのコスト要因よりも小さい、すなわち各ラベルなしデータ点の最終解に対する個々の寄与はラベル付きデータ点の個々の寄与よりも常に小さいからである。しかしながら、ラベルなしデータの量がラベル付きデータの数よりもはるかに大きい場合には、ラベルなしデータは依然として、ラベル付きデータよりも最終解に影響を与え得る。さらに、コスト要因のスケーリングと推定クラス確率を用いたラベルの事前確率の更新との結合は、上に概説したブリッジ効果の問題を解決する。最初のＭ段階で、ラベルなしデータは、極めて平坦な分類スコアの関数として期待ラベルをもたらす小さいコスト要因を有し（図１参照）、従って、小さい重みにすぎないが、ある程度まで、全ラベルなしデータは超平面上でプルすることが可能である。さらに、ラベルの事前確率の更新の結果として、分離超平面から離れたラベルなしデータはゼロに近い期待ラベルを割り当てられないが、数回の繰り返しの後に、ｙ＝＋１またはｙ＝−１に近いラベルが割り当てられ、かくして、ラベル付きデータのようにゆっくりと処理される。

Scaling the cost factor of data points by | <y> | alleviates the problem that unlabeled data may have a stronger cumulative pull on the hyperplane than labeled data. This is because the cost factor for unlabeled data is now smaller than the labeled cost factor, ie, the individual contribution to the final solution of each unlabeled data point is always smaller than the individual contribution of the labeled data point. However, if the amount of unlabeled data is much larger than the number of labeled data, unlabeled data can still affect the final solution more than labeled data. Further, the combination of cost factor scaling and label prior probability update using estimated class probabilities solves the bridge effect problem outlined above. In the first M stages, the unlabeled data has a small cost factor that yields the expected label as a function of a very flat classification score (see FIG. 1) and is therefore only a small weight, but to some extent, not all labels Data can be pulled on the hyperplane. Furthermore, as a result of updating the label prior probability, unlabeled data away from the separation hyperplane is not assigned an expected label close to zero, but after several iterations, it is close to y = + 1 or y = -1. Labels are assigned and thus processed slowly as labeled data.

  In a particular implementation of the method of an embodiment of the present invention, by assuming a Gaussian prior distribution with mean zero and unit variance for the decision function parameter Θ:

決定関数パラメータに対する事前分布は、当面の特定の分類上の問題に関する重要な従来知識を組み込んでいる。分類上の問題にとって重要な決定関数パラメータの他の事前分布は、例えば、多項分布、ポアソン分布、コーシー分布（Ｂｒｅｉｔ−Ｗｉｇｎｅｒ）、マクスウェル−ボルツマン分布、またはボーズ−アインシュタイン分布である。

Prior distributions for decision function parameters incorporate important prior knowledge about the particular classification problem at hand. Other prior distributions of decision function parameters that are important for classification problems are, for example, the multinomial distribution, Poisson distribution, Breit-Wigner, Maxwell-Boltzmann distribution, or Bose-Einstein distribution.

The prior distribution for decision function threshold b is given by a Gaussian distribution with mean μ _b and variance σ _b ² .

データ点の分類マージンγ_ｉの事前分布として、

As a prior distribution of data point classification margin γ _i ,

が選ばれ、ここで、ｃはコスト要因である。この事前分布は、式ｅｘｐ［−ｃ（ｌ−γ）］の形を有する本明細書において参照するＪａａｋｋｏｌａで用いられるものとは異なっている。式９で与えられた形が本明細書において参照するＪａａｋｋｏｌａで用いられる形を越えることが好ましく、なぜならば、Ｊａａｋｋｏｌａの形が１より小さいコスト要因に対してさえも正の期待マージンをもたらすのに対して、式ｅｘｐ［−ｃ（ｌ−γ）］は、ｃ＜１に対して負の期待マージンをもたらすからである。

Where c is a cost factor. This prior distribution is different from that used in Jaakcola referred to herein having the form exp [−c (l−γ)]. It is preferred that the form given by Equation 9 exceed the form used in Jaakcola referenced herein, because the Jaakola form provides a positive expectation margin even for cost factors less than 1. In contrast, the expression exp [−c (l−γ)] provides a negative expected margin for c <1.

  Given these prior distributions, it is easy to determine the corresponding partition function Ζ (see, eg, TM Cover and JA Thomas “Elements of Information Theory”, John Wiley & Sons, Inc.). See) (Cover), objective function

は、

Is

となる。本明細書において参照するＪａａｋｋｏｌａによれば、Ｍ段階の目的関数は、

It becomes. According to Jaakcola referred to in this specification, the M-stage objective function is

となり、Ｅ段階の目的関数は、

The objective function of stage E is

となる。ここで、ｓ_ｔは先のＭ段階で決定されたｔ番目のデータ点の分類スコアであり、ｐ_０，ｔ（ｙ_ｔ）はデータ点の２値ラベル事前確率である。ラベルの事前確率は、ラベル付きデータに対してはｐ_０，ｔ（ｙ_ｔ）＝１に、ラベルなしデータに対しては、ｐ_０，ｔ（ｙ_ｔ）＝１／２の情報を与えない事前確率またはクラスの事前確率に初期化される。

It becomes. Here, s _t is the classification score of the t-th data point determined in the previous M stage, and p _{0, t} (y _t ) is the binary label prior probability of the data point. The prior probability of the label does not give information of p _{0, t} (y _t ) = 1 for labeled data and p _{0, t} (y _t ) = 1/2 for unlabeled data. Initialized to prior probabilities or class prior probabilities.

  The section entitled M-stage in this specification describes an algorithm for solving the M-stage objective function. Also, the chapter entitled “E stage” in this specification describes the algorithm of the E stage.

  The EstimateClassProvability stage in row 5 of Table 3 uses training data to determine calibration parameters for changing the classification score to class membership probability, ie, the probability of the class given the score p (c | s). To do. A related method for estimating score calibration for probability is described in J. Org. Platt “Probabilistic outputs for support vector machines and comparison to legal methods,” pages 61-74, 2000 (Platt), and Zadrozny and C.I. Elkan "Transforming classifiers into accurate multi-class probabilities Estimates", 2002 (Zadrozny).

With particular reference to FIG. 3, a cross (×) indicates a negative example with a label, a plus sign (+) indicates a positive example with a label, and a circle (◯) indicates unlabeled data. The various plots show the separation hyperplane determined at various iteration times of the M stage. The 20th iteration shows the final solution chosen by the improved transductive MED classifier. FIG. 3 shows an improved transductive MED classification algorithm applied to the toy data set introduced above. The usage parameters are c = 10, σ _b ² = 1, and μ _b = 0. A different c results in a separation hyperplane located between x≈−0.5 and x = 0, so that for c <3.5, the hyperplane is the right of one unlabeled data with x <0 And c ≧ 3.5, to the left of this unlabeled data point.
With particular reference to FIG. 4, a control flow diagram illustrating a method for classifying unlabeled data according to one embodiment of the present invention is shown. The method 100 begins at step 102 and accesses stored data 106 at step 104. The data is stored in a storage area and includes labeled data, unlabeled data, and at least one predetermined cost factor. Data 106 includes data points having assigned labels. The assigned label identifies whether the labeled data point is intended to be included in a particular category or excluded from a particular category.

  Once the data is accessed at step 104, the method of one embodiment of the present invention then determines the prior probabilities of the data point labels using the data point label information at step 108. Then, at step 110, an expected label for the data point is determined according to the prior probability of the label. Along with the expected label determined in step 110, labeled data, unlabeled data, and cost factors, step 112 includes iterative training of the transductive MED classifier by scaling the cost factor unlabeled data points. During each iteration of the calculation, the cost factor for the data points is scaled. Thus, the MED classifier learns through iterative iterations of computation. The trained classifier then accesses the input data 114 at step 116. The trained classifier may then complete the input data classification step at step 118 and ends at step 120.

  It should be understood that 106 unlabeled data and input data 114 may be derived from a single source. Thus, input data / unlabeled data can be used in 112 iteration processes, which are then used to classify at 118. Further, one embodiment of the present invention includes a feedback mechanism for the input data 114 to provide the input data to the data stored at 106 so that 112 MED classifiers can receive new data input. It is intended to be able to learn dynamically from.

  With particular reference to FIG. 5, a control flow diagram illustrating another classification scheme for unlabeled data of one embodiment of the present invention, including user-defined prior probability information, is shown. Method 200 begins at step 202 and accesses data 206 stored at step 204. Data 206 includes labeled data, unlabeled data, predetermined cost factors, and prior probability information provided by the user. The labeled data 206 includes data points with assigned labels. The assigned label identifies whether the labeled data point is intended to be included in a particular category or excluded from a particular category.

  At step 208, an expected label is calculated from the 206 data. The expected label is then used at step 210 to iteratively train the transductive MED classifier along with labeled data, unlabeled data, and cost factors. 210 iterations scale the cost factor of unlabeled data at each computation point. The calculation continues until the classifier is properly trained.

  The trained classifier then accesses the input data from the input data 212 at 214. The trained classifier may then complete the step of classifying input data at step 216. As with the process and method described in FIG. 4, input data and unlabeled data may be derived from a single source and input to the system at both 206 and 212. Thus, the input data 212 can affect training at 210, and as a result, the process can dynamically change over time with continuous input data.

  In both methods shown in FIGS. 4 and 5, the monitor may determine whether the system has reached convergence. Convergence can be determined when the change in the hyperplane during each iteration of the MED calculation falls below a predetermined threshold. In an alternative embodiment of the invention, this threshold may be determined when the determined expected label change falls below a predetermined threshold. If convergence is reached, the iterative training process may end.

  With particular reference to FIG. 6, a more detailed control flow diagram of the iterative training process of at least one embodiment of the method of the present invention is shown. Process 300 begins at step 302, where data is accessed from data 306. Data 306 may include labeled data, unlabeled data, at least one predetermined cost factor, and prior probability information. The labeled data point 306 is a label that identifies whether the data is a training example for data points that should be included in the specified category or a training example for data points that should be excluded from the specified category. Including. The prior probability information 306 includes probability information of labeled data sets and unlabeled data sets.

  At step 308, an expected label is determined from the data from the prior probability information at 306. At step 310, the cost factor for each unlabeled data is scaled in proportion to the absolute value of the expected label for the data point. Then, at step 312, using the labeled and unlabeled data as training examples according to their expected labels, determining a decision function that maximizes the margin between the included training examples and the excluded training examples Trains the MED classifier. At step 314, a classification score is determined using the 312 trained classifier. At 316, the classification score is calibrated against the class membership probability. In step 318, the prior probability information of the label is updated based on the class membership probability. At step 320, a MED calculation is performed to determine the probability distribution of labels and margins. Here, the previously determined classification score is used for MED calculation. As a result, a new expected label is calculated in step 322, and the expected label is updated in step 324 using the calculation result from step 322. In step 326, the method determines whether convergence has been reached. If so, the method ends at step 328. If convergence has not been reached, another iteration of the method starting at step 310 is completed. The iteration is repeated until convergence is reached, so that the MED classifier is iteratively trained. Convergence can be reached when the change in the decision function during each iteration of the MED calculation falls below a predetermined value. In an alternative embodiment of the present invention, convergence may be reached when the determined change in expected label value falls below a predetermined threshold.

  FIG. 7 illustrates a network architecture 700 according to one embodiment. As shown, a plurality of remote networks 702 including a first remote network 704 and a second remote network 706 are provided. A gateway 707 may be coupled between the remote network 702 and the adjacent network 708. In the context of the present network architecture 700, each of the networks 704, 706 may take any form including, but not limited to, a LAN, a WAN such as the Internet, a PSTN, an internal telephone network, and the like.

  In use, the gateway 707 serves as an entry point from the remote network 702 to the adjacent network 708. Thus, the gateway 707 can function as a router that directs a given data packet that reaches the gateway 707 and as a switch that provides the actual path to and from the gateway 707 for the given packet.

  Further included is at least one data server 714 coupled to the neighboring network 708 and accessible from the remote network 702 via the gateway 707. It should be noted that the data server (s) 714 may include any type of computing device / groupware. A plurality of user devices 716 are coupled to each data server 714. Such user equipment 716 may include a desktop computer, laptop computer, handheld computer, printer, or any other type of logic. It should be noted that in one embodiment, user equipment 717 can also be directly coupled to any network.

  A facsimile machine 720 or a series of facsimile machines 720 may be coupled to one or more of the networks 704, 706, 708.

  It should be noted that the database and / or additional components can be used with or integrated with any type of network element coupled to the networks 704, 706, 708. Within the context of this description, a network element may refer to any component of the network.

  FIG. 8 illustrates an exemplary hardware environment associated with the user device 716 of FIG. 7, according to one embodiment. The figure shows a typical hardware configuration of a workstation having a central processing unit 810, such as a microprocessor, and a number of other units interconnected via a system bus 812.

  The workstation shown in FIG. 8 includes a random access memory (RAM) 814, a read only memory (ROM) 816, an I / O adapter 818 for connecting peripheral devices such as a magnetic disk device 820 to the bus 812, Communicates workstation with user interface adapter 822 for connecting keyboard 824, mouse 826, speaker 828, microphone 832, and / or other interface devices such as touch screens and digital cameras (not shown) to bus 812 A communication adapter 834 for connecting to a network 835 (eg, a data processing network) and a display adapter 836 for connecting the bus 812 to the display device 838 are included.

  With particular reference to FIG. 9, an apparatus 414 of one embodiment of the present invention is shown. One embodiment of the present invention includes a memory device 814 for storing labeled data 416. Each labeled data point 416 includes a label indicating whether the data point is an example training for a data point included in a specified category or an example training for a data point excluded from the specified category. Memory 814 also stores unlabeled data 418, prior probability data 420, and cost factor data 422.

  The processor 810 accesses data from the memory 814 and trains a binary classifier using transductive MED calculations so that it can classify unlabeled data. The processor 810 uses the cost factors and training examples from the labeled and unlabeled data and scales the cost factors as a function of the expected label value before being re-input to the processor 810. Iterative transductive computation is used by affecting 422 data. Thus, cost factor 422 changes with each iteration of MED classification by processor 810. Once processor 810 properly trains the MED classifier, the processor may then construct a classifier to classify unlabeled data into classified data 424.

Prior art transductive SVM and MED formulations provide an exponential increase in possible label assignments, and for practical use it is necessary to develop approximate equations. In an alternative embodiment of the present invention, a different formulation of transductive MED classification is introduced that does not increase the possible label assignments exponentially and allows general closed-form solutions. For a linear classifier, this problem is formulated as follows: The distribution p (Θ), bias distribution p (b), and data point classification margin p (γ) with respect to the hyperplane parameter are set to the minimum cullback for each prior distribution p ₀ to which these combined probability distributions are combined. Ask to have a Liver Divergence KL, ie

であり、ラベル付きデータに対して以下の制約に従い、

And follows the following restrictions for labeled data:

ラベルなしデータに対して以下の制約に従い、

Subject to the following restrictions for unlabeled data:

ここで、ΘＸ_ｔは、分離超平面の重みベクトルとｔ番目のデータ点の特徴ベクトルとのドット積である。ラベルに対する事前分布は必要ではない。ラベル付きデータは、それらの既知のラベルに従って分離超平面の右側に位置するように制約されているが、ラベルなしデータに対する唯一の要求条件は、超平面までのラベルなしデータの距離の２乗がマージンよりも大きいということである。要約すると、本発明のこの実施形態は、選択された事前分布に最も近く、ラベル付きデータを適切に分離し、かつマージン間にラベルなしデータを全く有しないという、妥協点となる分離超平面を求める。利点は、ラベルに対する事前分布を導入する必要がなく、従って、指数関数的に増加するラベル割り当てに関する問題が回避されることである。

Here, ΘX _t is the dot product of the weight vector of the separation hyperplane and the feature vector of the t th data point. Prior distribution for labels is not necessary. Although labeled data is constrained to lie to the right of the separation hyperplane according to their known labels, the only requirement for unlabeled data is that the square of the distance of unlabeled data to the hyperplane is It is larger than the margin. In summary, this embodiment of the present invention produces a compromise hyperplane that is closest to the selected prior distribution, properly separates the labeled data, and has no unlabeled data between the margins. Ask. The advantage is that there is no need to introduce a prior distribution for the labels, thus avoiding problems with exponentially increasing label allocation.

  In a specific implementation of another embodiment of the present invention, using the prior distributions given in Equation 7, Equation 8, and Equation 9 for hyperplane parameters, bias, and margin, the following distribution function is obtained: ,

ここで、ｔはラベル付きデータの添え字であり、ｔ´はラベルなしデータの添え字である。下記の表記法を用いると、

Here, t is a subscript of labeled data, and t ′ is a subscript of unlabeled data. Using the following notation:

式１６は、以下のように書き換えられる。

Equation 16 can be rewritten as follows:

積分の後に、以下の分配関数が得られる。

After integration, the following partition function is obtained:

すなわち、最終目的関数は、

That is, the final objective function is

となる。目的関数

It becomes. Objective function

は、本明細書においてＭ段階と題する章で述べられる、既知のラベルの場合の手法に類似した手法を適用することによって解かれ得る。差異は、最大マージン項の二次形式におけるマトリックスＧ_３ ^−１が、ここで非対角項を有する点である。

Can be solved by applying a technique similar to that of the known label case described herein in the section entitled M Stage. The difference is that the matrix G ₃ ⁻¹ in the quadratic form of the maximum margin term now has off-diagonal terms.

  In addition to classification, there are many applications of the method of the present invention that incorporate a framework of maximum entropy identification. For example, MED can be applied to solve general data classifications, any kind of discriminant functions and prior distributions, regression models and graphical models (T. Jebara “Machine Learning and Generalized”, Kluwer Academic Publishers). (Jebara).

  Applications of embodiments of the present invention can be formulated as purely recursive learning problems with known labels and as transductive learning problems with labeled and unlabeled training examples. In the latter case, improvements to the transductive MED classification algorithm described in Table 3 are similar for general transductive MED classification, transductive MED regression, and graphical model transductive MED learning. It is applicable to. Thus, for purposes of this disclosure and the claims, the term “classification” may include regression or graphical models.

(M stage)
According to Equation 11, the M-stage objective function is

となる。これにより、ラグランジュ乗数λ_ｔは、Ｊ_Ｍを最大化することによって決定される。

It becomes. Thereby, the Lagrange multiplier λ _t is determined by maximizing J _M.

Without the redundant constraint of λ _t <c, the Lagrangian for the above dual problem is

となる。最適性に対して必要かつ十分なＫＫＴ条件は、

It becomes. The necessary and sufficient KKT condition for optimality is

となる。ここで、Ｆ_ｔは、

It becomes. Where F _t is

である。最適点において、基底は期待バイアス

It is. At the optimal point, the basis is expected bias

と等しくなり、

Is equal to

が得られる。

Is obtained.

These equations can be summarized by considering two examples with the δ _t λ _t = 0 constraint. The first example is λ _t = 0 for all, and the second example is 0 <λ _t <c for all. Applied to the SVM algorithm. Keerthi, S .; Shevade, C.I. Bhattacharya, and K.A. A third example is not necessary, as described in Murthy “Improvements to plait's smo algorithm for svm classifier design”, 1999 (Keerthy). The potential function in this formulation maintains λ _t ≠ c.

最適条件に到達するまでに、一部のデータ点ｔに対するこれらの条件の侵害が存在する。すなわち、λ_ｔがゼロでないときにはＦ_ｔ≠−〈ｂ〉、またはλ_ｔがゼロのときにはＦ_ｔ〈ｙ_ｔ〉＜−〈ｂ〉〈ｙ_ｔ〉、である。残念なことに、〈ｂ〉の計算は、最適なλ_ｔのそれなくしては不可能である。これに対する良解は、以下の３つの組を構築することによって、再び本明細書において参照するＫｅｅｒｔｈｉから借用される。

By the time the optimal conditions are reached, there is a violation of these conditions for some data points t. That is, when the lambda _t is non-zero _{F t} ≠ - <b>, or lambda _t when the zero _{F t <y t><-} <b><yt>, is. Unfortunately, the calculation of <b> is not possible without that of the optimal λ _t . A good answer to this is borrowed from Keerthi, referred to herein again, by constructing the following three sets:

これらの組を利用して、以下の定義を用いた最も極端な最適性条件違反を定義することができる。Ｉ_０の要素は、それらが−〈ｂ〉に等しくないときは常に違反であり、従って、Ｉ_０からの最大Ｆ_ｔおよび最小Ｆ_ｔは、違反の候補である。Ｉ_１の要素は、Ｆ_ｔ＜−〈ｂ〉のときに違反であり、従って、Ｉ_１からの最小要素は、もしあるとすれば、最も極端な違反である。最後に、Ｉ_４の要素は、Ｆ_ｔ＞−〈ｂ〉のときに違反であり、それはＩ_４からの最大要素を違反候補にする。従って、−〈ｂ〉は以下に示すように、これらの組に関する最小および最大によって制限される。

These sets can be used to define the most extreme optimality condition violations using the following definitions: The elements of I ₀ are in violation whenever they are not equal to-<b>, so the maximum F _t and minimum F _t from I ₀ are candidates for violation. The element of I ₁ is violated when F _t <-<b>, so the smallest element from I ₁ is the most extreme violation, if any. Finally, the element of I ₄ is in violation when F _t > − <b>, which makes the largest element from I _{4 a} candidate for violation. Thus,-<b> is limited by the minimum and maximum for these sets, as shown below.

最適な−ｂ_ｕｐと−ｂ_ｌｏｗとは等しくなければならず、すなわち−〈ｂ〉であるので、−ｂ_ｕｐと−ｂ_ｌｏｗとの間のギャップを減らすことが、訓練アルゴリズムを収束に向けてプッシュする。さらに、ギャップはまた、数値的収束を判断するための手法として、測定され得る。

Optimal -b _up and -b _low must be equal, i.e.-<b>, so reducing the gap between -b _up and -b _low will make the training algorithm converge. To push. Further, the gap can also be measured as a technique for determining numerical convergence.

As described above, the value of b = <b> is unknown until convergence. The method of this alternative embodiment differs in that only one example can be optimized at a time. Thus, the training heuristic is to go back and forth between the I ₀ example and all examples every other time.

(E stage)
The objective function of the E stage of Equation 12 is

であり、ここでｓ_ｔは、先のＭ段階で決定されたｔ番目のデータ点の分類スコアである。ラグランジュ乗数λ_ｔは、

Where s _t is the classification score of the t th data point determined in the previous M stages. The Lagrange multiplier λ _t is

を最大化することによって決定される。

Is determined by maximizing.

Without the redundant constraint of λ _t <c, the Lagrangian for the above dual problem is

となる。最適性に対して必要かつ十分なＫＫＴ条件は、

It becomes. The necessary and sufficient KKT condition for optimality is

である。ＫＫＴ条件を最適化することによってラグランジュ５乗数に対する解を求めることは、ＫＫＴ条件が例を分解する（ｆａｃｔｏｒｉｚｅ）ので、例を１回パスすることによって行われ得る。

It is. Finding a solution for the Lagrangian multiplier by optimizing the KKT condition can be done by passing the example once, as the KKT condition factorizes the example.

For the labeled example, the expected label <y _t > has P _{0, t} (y _t ) = 1 and P _{0, t} (−y _t ) = 0, and the KKT condition

に簡略化し、ラベル付き例のラグランジュ乗数に対する解として、

As a solution to the Lagrangian multiplier in the labeled example,

をもたらす。ラベルなし例に対して、式３５は解析的に解くことはできないが、しかしながら、式３５を満たす各ラベルなし例のラグランジュ乗数に対して、例えば線形探索を適用することによって、決定されねばならない。

Bring. For the unlabeled example, Equation 35 cannot be solved analytically, however, it must be determined, for example, by applying a linear search to the Lagrangian multiplier for each unlabeled example that satisfies Equation 35.

  The following are some non-limiting examples made possible by the techniques described above, their derivatives or variations, and other techniques known in the art. Each example includes a suitable algorithm and optional algorithms or parameters that can be implemented in a basic preferred approach.

  In one embodiment presented in FIG. 10, labeled data points are received at step 1002, where the data points are training examples for data points to be included in a specified category, or a specified category. Each labeled data point has at least one label that indicates whether it is a training example for a data point excluded from. Further, unlabeled data points are received at step 1004 along with at least one predetermined cost factor of labeled and unlabeled data points. Data points can include any medium, such as words, images, sounds, and the like. Prior probability information for labeled and unlabeled data points may also be received. Also, the included training example labels may be mapped to a first numerical value, such as +1, and the excluded training example labels may be mapped to a second numerical value, such as -1. Further, at least one predetermined cost factor of labeled data points, unlabeled data points, input data points, labeled data points and unlabeled data points may be stored in the memory of the computer.

  Further, at step 1006, the transductive MED classifier is trained by iterative calculation using the at least one cost factor described above and labeled and unlabeled data points as training examples. For each iteration of the calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, for example, the absolute value of the expected label of the data point, and the prior probability of the data point label is the data point class It is adjusted based on the estimated probability of belonging, thereby ensuring stability. Also, the transductive classifier can learn using prior probability information of labeled and unlabeled data, which further improves stability. The iterative step of training the transductive classifier is that the change in the expected expected label value is determined until the data value reaches convergence, eg, when the change in the decision function of the transductive classifier is below a predetermined threshold. When it falls below a predetermined threshold, etc. can be repeated.

  Further, at step 1008, the trained classifier is applied to classify at least one of unlabeled data points, labeled data points, and input data points. Input data points may be received before or after the classifier is trained, or may not be received at all. A decision function that minimizes KL divergence for the prior probability distribution of decision function parameters given included and excluded training examples also learns labeled and unlabeled data points according to their expected labels It can be determined using as an example. Alternatively, the decision function can be determined by minimum KL divergence using a multinomial distribution for the decision function parameters.

  At step 1010, the classification of the classified data points, or a derivative thereof, is output to at least one of the user, another system, and another process. The system can be remote or local. Examples of categorical derivatives may include, but are not limited to, classified data points themselves, representations of classified data points or their identifiers, or host files / documents.

  In another embodiment, computer executable program code is deployed to and executed on a computer system. The program code includes instructions for accessing labeled data points stored in the memory of the computer, each labeled data point being trained on a data point to be included in the category in which the data point is specified. It has at least one label indicating whether it is an example or a training example for data points excluded from a specified category. In addition, the computer code includes instructions for accessing unlabeled data points from the computer memory, and instructions for accessing at least one predetermined cost factor of labeled and unlabeled data points from the computer memory. Including. Prior probabilities information for labeled and unlabeled data points stored in the computer's memory can also be accessed. Also, the included training example labels may be mapped to a first numerical value, such as +1, and the excluded training example labels may be mapped to a second numerical value, such as -1.

  Further, the program code trains the transductive classifier by iterative calculations using at least one stored cost factor and stored labeled data points and stored unlabeled data points and training examples. Instructions are provided. Also, for each iteration of the calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value of the data point, eg, the absolute value of the expected label of the data point. Also, prior probability information can be adjusted for each iteration based on the estimated value of class membership probability of the data points. The iterative step of training the transductive classifier is that until the convergence of the data value is reached, for example, when the change of the decision function of the transductive classifier falls below a predetermined threshold, the change in the determined expected label value is When it falls below a predetermined threshold, etc. can be repeated.

  Further, the program code applies a trained classifier to classify at least one of the unlabeled data points, the labeled data points, and the input data points, and the classified data points. Instructions for outputting the classification, or a derivative thereof, to a user, another system, and another process. Also, given the included training examples and the excluded training examples, the decision function that minimizes KL divergence for the prior probability distribution of the decision function parameters is to use labeled and unlabeled data points as their expected labels. As a learning example and can be determined.

  In yet another embodiment, the data processing apparatus is (i) a training example for data points that are included in a specified category or a training example for data points that are excluded from the specified category. To store labeled data points each having at least one label, (ii) unlabeled data points, and (iii) at least one predetermined cost factor of labeled and unlabeled data points At least one memory. This memory may also store prior probability information for labeled and unlabeled data points. Also, the included training example labels may be mapped to a first numerical value, such as +1, and the excluded training example labels may be mapped to a second numerical value, such as -1.

  In addition, the data processing apparatus uses transductive maximum entropy identification (MED), using at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples. And a transductive classifier training apparatus for repeatedly teaching the transductive classifier. Further, at each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value of the data point, eg, the absolute value of the expected label of the data point. Also, in each iteration of the MED calculation, the prior probability information can be adjusted based on an estimate of the class membership probability of the data points. For example, when the change in the decision function of the calculation of the transductive classifier falls below a predetermined threshold, or when the change in the expected expected label value falls below the predetermined threshold, the apparatus converges the data value. And a means for terminating the calculation simultaneously with the determination of convergence.

  Further, using the trained classifier, at least one of unlabeled data points, labeled data points, and input data points are classified. Furthermore, given the included training examples and the excluded training examples, the decision function that minimizes the KL divergence for the prior probability distribution of the decision function parameters is able to reduce the labeled and unlabeled data according to their expected labels. It can be determined by the processor using as a learning example. Also, the classification of the classified data points, or a derivative thereof, is output to at least one of the user, another system, and another process.

  In a further embodiment, the product comprises a computer readable program storage medium, which unambiguously embodies one or more programs of instructions executable by a computer to perform data classification. In use, each has at least one label indicating whether the data points are training examples for data points included in the specified category or are training examples for data points excluded from the specified category A labeled data point is received. Further, unlabeled data points and at least one predetermined cost factor of labeled data points and unlabeled data points are received. Prior probability information for labeled and unlabeled data points can also be stored in the memory of the computer. Also, the included training example labels may be mapped to a first numerical value, such as +1, and the excluded training example labels may be mapped to a second numerical value, such as -1.

  Further, the transductive classifier is trained by iterative maximum entropy identification (MED) computation using at least one stored cost factor and stored labeled and unlabeled data points as training examples. In each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value of the data point, eg, the absolute value of the expected label of the data point. Also, in each iteration of MED, prior probability information can be adjusted based on an estimate of the class membership probability of the data points. The iterative step of training the transductive classifier is that until the convergence of the data value is reached, for example, when the change of the decision function of the transductive classifier falls below a predetermined threshold, the change in the determined expected label value is It can be repeated until, for example, when it falls below a predetermined threshold.

  Further, the input data points are accessed from a computer memory and a trained classifier is applied to classify at least one of the unlabeled data points, the labeled data points, and the input data points. Also, given the included training examples and the excluded training examples, a decision function that minimizes KL divergence for the prior probability distribution of decision function parameters is used for labeling and unlabeled data according to their expected labels. It can be determined using as a learning example. Further, the classification of the classified data points, or a derivative thereof, is output to at least one of the user, another system, and another process.

  In yet another embodiment, a method for classifying unlabeled data in a computer-based system is presented. In use, each has at least one label that indicates whether the data points are training examples for data points that should be included in the specified category or are training examples for data points excluded from the specified category. A labeled data point having is received.

  Further, labeled data points and unlabeled data points are received, as well as pre-label probability information for labeled and unlabeled data points. Further, at least one predetermined cost factor is received for labeled data points and unlabeled data points.

Further, an expected label for each labeled and unlabeled data point is determined based on the prior probability of the data point label. The following substeps are repeated until the data values substantially converge. That is,
Produce a scaled cost value for each unlabeled data point in proportion to the absolute value of the expected label for the data point,
Use labeled and unlabeled data as training examples according to their expected labels to minimize KL divergence for prior probability distributions of decision function parameters given included training examples and excluded training examples Train a maximum entropy discriminator (MED) classifier by determining a decision function to
Use a trained classifier to determine classification scores for labeled and unlabeled data points;
Calibrate the output of the trained classifier to the class membership probability,
Update the prior probabilities for unlabeled data points according to the determined class membership probabilities,
Determining the probability distribution of labels and margins using maximum entropy identification (MED) using the updated prior probabilities and previously determined classification scores;
・ A new expected label is calculated using the probability distribution of the previously determined label,
Update the expected label for each data point by interpolating the new expected label with the expected label from the previous iteration.

  Also, the classification of the input data points or derivatives thereof is output to at least one of the user, another system, and another process.

  Convergence can be reached when the change in the decision function falls below a predetermined threshold. Furthermore, convergence can also be reached when the calculated change in expected label value falls below a predetermined threshold. Further, the included training example labels may have any value, eg, a value of +1, and the excluded training example labels may have any value, eg, a value of −1.

  A method for classifying documents in one embodiment of the present invention is presented in FIG. In use, at step 1100, at least one seed document having a known confidence level is received along with an unlabeled document and at least one predetermined cost factor. Seed documents and other items may be received from computer memory, users, network connections, etc., and may be received after a request from a system executing the method. The at least one seed document may have a label that indicates whether the document is included in a specified category, may include a list of keywords, or any other attribute that may assist in document classification. Can have. Further, in step 1102, the transductive classifier is trained by iterative calculations using at least one predetermined cost factor, at least one seed document, and unlabeled documents, where for each iteration of the calculation: Cost factors are adjusted as a function of expected label values. Data point label prior probabilities for labeled and unlabeled documents can also be received, where for each iteration of the calculation, the data point label prior probabilities are adjusted according to the data point class membership probability estimate. Can be done.

  Further, after at least some iterations, at step 1104, a confidence score for the unlabeled document is stored, and at step 1106, the identifier of the unlabeled document with the highest confidence score is determined by the user, another system, and another Output to at least one of the processes. Identifiers can be electronic copies of the documents themselves, their parts, their titles, their names, their file names, pointers to their documents, and so on. The confidence score can also be stored after each iteration, in which case the identifier of the unlabeled document with the highest confidence score is output after each iteration.

  One embodiment of the present invention can find patterns that link the first document to the remaining documents. The work of disclosure procedures is an area where this pattern discovery is particularly valuable. For example, pre-trial legal disclosure procedures require the investigation of large volumes of documents that may be relevant to the current litigation. The ultimate goal is to discover “definitive evidence”. In another example, the day-to-day work of inventors, patent examiners, and patent attorneys is to assess the novelty of technology through prior art searches. Specifically, the task is to search all published patents and other public relations to find documents in this set that may be relevant to the particular technology being examined for novelty.

  The job of the disclosure procedure involves the discovery of a document or a set of documents in a set of data. Having obtained the initial document or concept, the user may desire to find a document associated with the initial document or concept. However, the view of the relationship between the initial document or concept and the target document, ie the document to be discovered, is only fully understood after the discovery has occurred. By learning from labeled data points and unlabeled documents, concepts, etc., the present invention can learn patterns and relationships between the initial document or documents and the target document.

  A method for analyzing documents associated with legal disclosure procedures in another embodiment of the present invention is presented in FIG. In use, at step 1200, a document associated with a legal matter is received. Such documents may include electronic copies of the documents themselves, their parts, their titles, their names, their file names, pointers to documents, etc. Further, at step 1202, a document classification technique is performed on the document. Further, at step 1204, an identifier of at least a portion of the document is output based on the document classification. Optionally, one that displays links between documents can be output.

  Document classification techniques may include any type of processing, such as transductive processing. For example, any inductive or transductive technique described above can be used. In a preferred approach, the transductive classifier is trained by iterative calculations using at least one predetermined cost factor, at least one seed document, and documents associated with legal matters. For each iteration of the calculation, the cost factor is preferably adjusted as a function of the expected label value, and this trained classifier is used to classify the received document. The process may further include receiving a prior probability of the data point label for the labeled and unlabeled documents, where for each iteration of the calculation, the prior probability of the data point label is the data point's It is adjusted according to the estimation of class membership probability. Further, the document classification technique may include one or more of support vector machine processing and maximum entropy identification processing.

  In yet another embodiment, a method for analyzing a prior art document is presented in FIG. In use, at step 1300, the classifier is trained based on the search query. At step 1302, a plurality of prior art documents are accessed. Such prior art documents may include any information published in any form prior to a given date. Such prior art may additionally or alternatively include any information that has not been published in any way at any time prior to a given date. Exemplary prior art documents can be any type of document, for example, a patent office publication, data retrieved from a database, a collection of prior art, a portion of a website, and the like. A document classification technique is also performed at step 1304 on at least a portion of the prior art document using a classifier, and at least a portion of the identifier of the prior art document is output at step 1306 based on the prior art document classification. Is done. This document classification technique may include any one or more processes, including support vector machine processing, maximum entropy identification processing, or any recursive or transductive technique described above. Alternatively, or as an alternative, what displays a link between documents can be output. In yet another embodiment, relevance scores for at least some conventional documents are output based on the classification of the documents.

  The search query may include at least a portion of patent information disclosure. Exemplary patent information disclosures include disclosures made by the inventor, patent provisional applications, non-provisional patent applications, foreign patent applications, or patent applications, etc. that summarize the invention.

  In one preferred approach, the search query includes at least a portion of a claim retrieved from a patent document or patent application document. In another approach, the search query includes at least a portion of a patent document or a summary of patent application documents. In yet another approach, the search query includes at least a portion of a summary retrieved from a patent document or patent application document.

  FIG. 27 illustrates a method for matching a document with a claim. At step 2700, the classifier is trained based on at least one claim in the patent document or patent application document. Accordingly, one or more claims or portions thereof may be used to train a classifier. In step 2702, a plurality of documents are accessed. Such documents may include prior art documents, documents describing products that are potentially infringing or withdrawing, and the like. In step 2704, a document classification technique is performed on at least some documents using a classifier. At step 2706, at least some document identifiers are output based on the document classification. A relevance score for at least some documents may also be output based on the classification of the documents.

  One embodiment of the present invention can be used to classify patent applications. For example, in the United States, patents and patent applications are currently classified by subject matter using the US Patent Classification (USPC) system. This task is currently done manually, and is therefore very expensive and time consuming. Such manual classification also suffers human error. The ability to classify patent documents or patent application documents into multiple classes increases the degree of complexity of such work.

  FIG. 28 illustrates a method for classifying patent applications according to one embodiment. At step 2800, a classifier is trained based on a plurality of documents known to be in a particular patent classification. Such documents can generally be patent documents and patent application documents (or parts thereof), but can also be summary votes describing the target subject matter of a particular patent classification. At step 2802, at least a portion of a patent document or patent application document is received. Some of this may include claims, summaries, abstracts, specifications, titles, and the like. At step 2804, a document classification technique is performed on at least a portion of the patent document or patent application document using the classifier. At step 2806, the classification of the patent document or patent application document is output. Optionally, the user can manually verify some or all classifications of the patent application.

  The document classification technique is preferably a yes / no classification technique. In other words, if the probability that a document is in a particular class is above the threshold, the determination is “yes” and the document belongs to this class. If the probability that the document is in a particular class is below the threshold, the determination is “no” and the document does not belong to this class.

  FIG. 29 illustrates yet another method for classifying patent applications. In step 2900, a document classification technique is performed on at least a portion of a patent document or patent application document using a classifier trained based on at least one document associated with a particular patent classification. Again, the classification technique is preferably a yes / no classification technique. In step 2902, the classification of the patent document or patent application document is output.

  In either of the methods shown in FIGS. 28 and 29, each method can be repeated using different classifiers trained on multiple documents known to fall into different patent classifications.

  Officially, patent classification should be based on the claims. However, it may also be desirable to perform a match between (any IP related content) and (any IP related content). In one example, one approach is to train using a patent specification and classify the application based on the patent claims. Another approach is to train using the description and claims and classify based on the summary. A particularly preferred approach is that any part of a patent document or patent application document is trained, but the same type of content is used during classification, i.e. the system trains based on the claims. If so, the classification is based on the claims. Document classification techniques may include any type of processing, such as transductive processing. For example, any inductive or transductive technique described above can be used. In a preferred approach, the classifier can be a transductive classifier, and the transductive classifier can be trained by iterative calculations using at least one predetermined cost factor, at least one seed document, and prior art documents. Here, for each iteration of the calculation, the cost factor is adjusted as a function of the expected label value, and a trained classifier can be used to classify the prior art document. Data point label prior probabilities for the seed document and the prior art document may also be received, where for each iteration of the calculation, the data point label prior probability is an estimate of the data point class membership probability. Can be adjusted accordingly. The seed document can be any document, such as a patent office publication, data retrieved from a database, a collection of prior art, a website, patent disclosure information, and the like.

  FIG. 14 illustrates one embodiment of the present invention in one approach. In step 1401, a set of data is read. It is desirable to find documents associated with the user in this set. In step 1402, the first seed document or seed documents are labeled. The document can be any type of document, such as a patent office publication, data retrieved from a database, a collection of prior art, a website, and the like. It is also possible to seed the transduction process with a different set of keywords or documents provided by the user. At step 1406, training of the transductive classifier is performed using labeled data and a given set of unlabeled data sets. At each label induction step during the iterative transduction process, the confidence score determined during label induction is stored. Once training is complete, documents that have achieved a high confidence score in the label induction step are displayed to the user in step 1408. Those documents with a high confidence score represent documents that are relevant to the user for discovery purposes. The display can be made in time order of the label induction step, starting with the first seed document and ending with the final set of documents found in the last label induction step.

  Another embodiment of the invention includes data organization and accurate classification, for example, coupled with business process automation. Arrangement and classification techniques may include any type of processing, such as transductive processing. For example, any inductive or transductive technique described above can be used. In the preferred approach, the entry key to the database is used as a label associated with some confidence level, depending on the expected cleanliness of the database. Then, a label that has an associated confidence level, ie, an expected label, is used to train the transductive classifier, which modifies the label (key) and uses the data in the database. Achieve consistent organization. For example, to allow automatic data extraction, eg determination of total price, order number, product quantity, shipping address, etc., the invoice must first be classified according to the company or individual that issued the invoice. is there. Typically, training examples are required to prepare an automatic classification system. However, examples of training provided by customers often include misclassified documents or other noise--for example, the cover of a fax--that are identified prior to training an automatic classification system to obtain an accurate classification. Must be removed. In another example, it helps to detect discrepancies between reports written by doctors and diagnoses in the field of patient records.

  In another example, it is known that the Patent Office is continuously carrying out a reclassification process, in which case the Patent Office (1) evaluates the existing branch of the Patent Office's taxonomy for confusion. , (2) restructuring the taxonomy to evenly distribute over-congested nodes, and (3) reclassify existing patents into the new structure. The transductive learning method presented here re-evaluates its taxonomy by the JPO and companies that outsource this work, and (1) a new taxonomy for a given major category. And (2) can be used to assist with reclassifying existing patents.

  Transduction learns from labeled and unlabeled data, thereby smoothing the transition from labeled data to unlabeled data. At one end of the spectrum is labeled data with complete preliminary knowledge. That is, the given label is correct without exception. At the other end is unlabeled data for which no prior knowledge is given. The organized data containing a certain level of noise constitutes mislabeled data and is located somewhere on the spectrum between the two extremes mentioned above. The labels given by the organization of the data can be trusted to some extent as correct, but not completely. Thus, transduction replaces existing data organization by assuming a certain level of error within a given organization of data, and interpreting these as uncertainty in the preliminary knowledge about label assignment. Can be used for organizing.

  A method for organizing data in one embodiment is presented in FIG. In use, at step 1500, a plurality of labeled data items are received, and at step 1502, a subset of data items for each of the plurality of categories is selected. Further, at step 1504, the uncertainty for data items in each subset is set to approximately zero, and at step 1506, the uncertainty for data items not present in the subset is set to a predetermined value that is not approximately zero. . Further, at step 1508, the transductive classifier is trained by iterative calculation using the uncertainty, data items in the subset, and data items that do not exist in the subset as training examples, and trained in step 1510. A classifier is applied to each labeled data item to classify each data item. Also, at step 1512, the classification of the input data item or derivative thereof is output to at least one of the user, another system, and another process.

  Further, the subset can be selected randomly and can be selected and verified by the user. The labels of at least some data items can be changed based on the classification. In addition, after the data item is classified, an identifier of the data item having a confidence level below a predetermined threshold can be output to the user. The identifier can be an electronic copy of the document itself, their parts, their title, their name, their file name, a pointer to the document, and so on.

  In one embodiment of the invention, as shown in FIG. 16, at step 1600, the user is presented with two choices for initiating the organizing process. One option is fully automatic organization at step 1602, where a specific number of documents are randomly selected and correctly organized for each concept or category. Alternatively, at step 1604, some documents are flagged for manual review and verification that one or more label assignments for each concept or category are properly organized. obtain. At step 1606, an estimate of the noise level in the data is received. At step 1610, the transductive classifier is trained using the verified (manually verified or randomly selected) data and the unverified data from step 1608. Once training is complete, the document is reorganized according to the new label. At step 1612, documents with low confidence below a certain threshold in label assignment are displayed to the user for manual review. At step 1614, documents having confidence levels above a certain threshold in label assignment are automatically modified according to the transductive label assignment.

  In another embodiment, a method for managing medical records is shown in FIG. In use, at step 1700, the classifier is trained based on the medical diagnosis, and at step 1702, multiple medical records are accessed. Further, at step 1704, a document classification technique is performed on the medical record using a classifier, and at step 1706, an identifier of at least one medical record having a low probability associated with a medical diagnosis is output. The document classification technique may include any type of processing, such as transductive processing, and includes one or more of any of the above described recursive techniques or transformers including support vector machine processing, maximum entropy identification processing, etc. It may include a ductive approach.

  In one embodiment, the classifier can be a transductive classifier, the transductive classifier can be trained by iterative calculations using at least one predetermined cost factor, at least one seed document, and medical records; Here, for each iteration of the calculation, the cost factor is adjusted as a function of the expected label value, and then a trained classifier can be used to classify the medical records. Prior probabilities of data point labels for seed documents and medical records may also be received, where for each iteration of the calculation, the prior probabilities of data point labels are in accordance with an estimate of the class membership probability of the data points. Can be adjusted.

  Another embodiment of the present invention is responsible for dynamic, shifting classification concepts. For example, in a format for processing an application, the document is classified using document layout information and / or content information to classify the document for subsequent processing. In many applications, documents are not static and evolve over time. For example, the content and / or layout of a document may change due to new legislation. Transductive classification automatically adapts to these changes and provides the same or equivalent classification accuracy despite the drifting classification concept. This is in contrast to rule-based systems or inductive classification methods that suffer from classification accuracy from the outset due to conceptual drift without manual adjustment. One example of this is invoice processing, which traditionally involves inductive learning or uses a rule-based system that utilizes invoice layout. Under these conventional systems, if layout changes occur, the system must be manually reconfigured by labeling new training data or defining new rules. However, the use of transduction eliminates manual reconfiguration by automatically adapting to small changes in the invoice layout. In another example, transductive classification can be applied to customer complaint analysis to monitor changes in the nature of such complaints. For example, a company may automatically combine product changes with customer complaints.

  Transduction can also be used to classify news articles. For example, news articles about the fight against terrorism, which began with an article on attacks by terrorists on September 11, 2001, went through the war in Afghanistan, and covered the news coverage of today's situation in Iraq. Specific.

  In yet another example, the taxonomy of an organism (alpha taxonomy) can change over time with evolution due to the creation of new species of organisms and the extinction of other species. These laws of classification system or taxonomy can be dynamic, with classification concepts shifting or changing over time.

By using input data to be classified as unlabeled data, transduction can recognize shifting classification concepts and thus dynamically adapt to evolving classification schemes. For example, FIG. 18 shows one embodiment of the present invention using transduction given the drifting classification concept. As shown in step 1802, the document set D _i enters the system at time t _i . At step 1804, the transductive classifier C _i is trained with the labeled and unlabeled data accumulated so far, and at step 1806, the documents in the set D _i are classified. If manual mode is used, documents having confidence levels below the user-specified threshold determined at step 1808 are presented to the user for manual review at step 1810. As shown in step 1812, in automatic mode, a document with a certain confidence level triggers the creation of a new category to be added to the system, which is then assigned to that new category. Documents having a confidence level above the selected threshold are classified into current categories 1 through N in steps 1820A-1820B. All documents in the current category that have been classified in the current category before step t _i are reclassified by the classifier C _{i in} step 1822 and all documents that are not classified in the previously assigned category are Steps 1824 and 1826 move to a new category.

  In yet another embodiment, a method for accommodating document content shifting is presented in FIG. The content of the document may include, but is not limited to, graphical content, character content, layout, numbering, and the like. Examples of shifts may include temporal shifts, style shifts (when two or more people work on one or more documents), processing shifts applied, layout shifts, and the like. At step 1900, at least one labeled seed document is received along with an unlabeled document and at least one predetermined cost factor. Documents may include, but are not limited to, customer complaints, invoices, form documents, receipts, and the like. Further, at step 1902, the transductive classifier is trained using at least one predetermined cost factor, at least one seed document, and an unlabeled document. Also, in step 1904, unlabeled documents having a confidence level that exceeds a predetermined threshold are classified into a plurality of categories using a classifier, and at least some categorized documents use the classifier in step 1906. Reclassified into categories. Further, at step 1908, the categorized document identifier is output to at least one of the user, another system, and another process. The identifier can be an electronic copy of the document itself, their parts, their title, their name, their file name, a pointer to the document, and so on. Furthermore, product changes can be combined with customer complaints and the like.

  In addition, unlabeled documents having a confidence level below a predetermined threshold can be moved to one or more new categories. Also, the transductive classifier can be trained by iterative calculations using at least one predetermined cost factor, at least one seed document, and an unlabeled document, where for each iteration of the calculation, the cost factor is It can be adjusted as a function of the expected label value and the unlabeled document can be classified using a trained classifier. Further, prior probabilities of data point labels for seed documents and unlabeled documents may be received, where for each iteration of the calculation, the prior probabilities of data point labels are estimates of class membership probabilities of data points. It can be adjusted according to the value.

  In another embodiment, a method for adapting patent classification to document content shifts is presented in FIG. At step 2000, at least one labeled seed document is received along with the unlabeled document. Unlabeled documents may include any type of document, such as patent application documents, court filing documents, information disclosure forms, document modifications, and the like. Seed document (s) may include patent document (s), patent application document (s), and the like. In step 2002, a transductive classifier is trained using at least one seed document and an unlabeled document, and an unlabeled document having a confidence level above a predetermined threshold is converted into a plurality of existing categories using the classifier. are categorized. The classifier can be any type of classifier, such as a transductive classifier, and the document classification technique can be any technique, such as support vector machine processing, maximum entropy identification processing, and the like. For example, any inductive or transductive technique described above can be used.

  Also, in step 2004, unlabeled documents having a confidence level below a predetermined threshold are classified into at least one new category using a classifier, and in step 2006, at least some categorized documents are classified. Reclassify into an existing category and at least one new category using the instrument. Further, at step 2008, the categorized document identifier is output to at least one of the user, another system, and another process. The transductive classifier can also be trained by iterative calculations using at least one predetermined cost factor, a search query, and a document, where for each iteration of the cost, the cost factor is a function of the expected label value. And a trained classifier can be used to classify the document. In addition, prior probabilities of data point labels for the search query and document may be received, and for each iteration of the calculation, the prior probabilities of the data point labels are adjusted according to an estimate of the class membership probability of the data points.

  Yet another embodiment of the invention is responsible for document drift in the field of document separation. One practical example for document separation involves the processing of mortgage documents. A series of various loan documents, such as a loan application, loan approval, loan request, loan amount, etc., is scanned for a loan-related document folder, and the various documents in the series of images are processed before further processing. Need to be confirmed. The document used is not static and can change over time. For example, the tax return form used in the loan related document folder may change over time due to changes in the law.

  Document separation solves the problem of finding document or partial document boundaries in a series of images. Common examples for generating a series of images are a digital scanner or a multifunction peripheral (MFP). As with classification, transduction can be used for document separation to deal with drift over time of documents and their boundaries. Static separation systems, such as rule-based systems or systems based on recursive learning solutions, cannot automatically adapt to drifting separation concepts. The performance of these static separation systems decreases with time whenever drift occurs. To maintain performance at its initial level, manually adapt the rules (for rule-based systems) or manually label new documents and retrain the system (inductive learning solution) In the case of). Both methods are time consuming and expensive. By applying transduction to document separation, it is possible to develop a system that automatically adapts to the separation concept drift.

  A method for document separation in one embodiment is presented in FIG. At step 2100, labeled data is received, and at step 2102 a series of unlabeled documents is received. Such data and documents may include statutory disclosure documents, notices of reasons for refusal, web page data, round trip letters between agents and clients, and the like. Further, at step 2104, probabilistic classification rules are adapted using transduction based on the labeled data and unlabeled documents, and at step 2106, the weights used for document separation are updated according to the probabilistic classification rules. The Also, at step 2108, a separation location within the series of documents is determined, and at step 2110, the indicator of the determined separation location within the series of documents is at least one of a user, another system, and another process. Are output. The sign may be an electronic copy of the document itself, their parts, their title, their name, their file name, pointer to the document, and so on. Further, at step 2112, the document is flagged with a code that correlates with the sign.

  FIG. 22 shows an implementation of the classification method and apparatus of the present invention used in conjunction with document separation. Automatic document separation is used to reduce manual effort involved in document separation and identification after digital scanning. One such document classification method uses the classification methods described herein, and combines the classification rules and uses an inference algorithm that subtracts the most likely separation from all available information, Automatically separate page sequences. In one embodiment of the present invention shown in FIG. 22, the transductive MED classification method of the present invention is employed for document separation. More specifically, a document page 2200 is inserted into a digital scanner 2202 or MFP and converted into a series of digital images 2204. The document page can be any type of document, for example, a page from a patent office publication, data retrieved from a database, a collection of prior art, a website, and the like. At step 2206, a series of digital images is input and the transduction is used to dynamically adapt the probabilistic classification rules. Step 2206 uses a series of images 2204 and unlabeled data 2208 as unlabeled data. In step 2210, the weights in the probabilistic network are updated and used for automatic document separation according to dynamically adapted classification rules. Output step 2212 is a dynamic adaptation of automatic insertion of separated images in which a series of digitized pages 2214 are interleaved with the automatic image of separating sheets 2216. In step 2212, images of the separating sheets are converted into a series of images. Insert automatically. In one embodiment of the present invention, software-generated separation page 2216 may also indicate the type of document that immediately follows or precedes separation page 2216. The system described here automatically adapts to the document's drifting separation concept that occurs over time and reduces the separation accuracy, like static systems such as rule-based solutions or recursive machine learning-based solutions. Will not suffer. A common example of a drifting separation or classification concept in the form of application processing is the revision of documents due to the enactment of new laws, as mentioned above.

  Furthermore, the system shown in FIG. 22 can be modified to the system shown in FIG. In the system shown in FIG. 23, a page 2300 is inserted into a digital scanner 2302 or MFP and converted into a series of digital images 2304. At step 2306, a series of digital images is input to dynamically adapt the probabilistic classification rules using transduction. Step 2306 uses a series of images 2304 as unlabeled data and labeled data 2308. Step 2310 updates the weights in the probabilistic network that are used for automatic document separation according to the adopted dynamically adapted classification rules. In step 2312, as described with reference to FIG. 18, the separation sheet image is not inserted, and in step 2312, the automatic insertion of separation information is dynamically adapted to set a coded description flag on the document image 2314. . In this way, the document page image can be entered into the visualized database 2316 and the document can be accessed by the software identifier.

  Yet another embodiment of the present invention can perform face recognition using transduction. As mentioned above, the use of transduction has many advantages, such as the relatively small number of training examples required, the availability of unlabeled examples for training, and the like. By taking advantage of the advantages described above, transductive face recognition can be implemented for crime arrest.

  For example, the Department of Homeland Security must ensure that terrorists are not allowed to board commercial aircraft. Part of the airport screening process may be to take a picture of each passenger at an airport checkpoint and attempt to recognize the person. The system can first be trained using a small number of examples from the limited photos available for suspected terrorists. There may also be more unlabeled photos of the same terrorists available that can also be used for training in other investigative authorities' databases. Thus, the transductive training device not only uses the initial sparse data to generate a functional face recognition system, but also uses unlabeled examples from other sources to improve performance. After processing the photos taken at the airport checkpoint, the transductive system can recognize the person in question more accurately than an inductive system that can be contrasted.

  In yet another embodiment, a face recognition method is presented in FIG. At step 2400, a seed image of at least one labeled face having a known confidence level is received. The at least one seed image may have a label that indicates whether the image is included in a specified category. Further, at step 2400, an unlabeled image is received from, for example, a police, government agency, lost child database, airport security department, or any other location, and at least one predetermined cost factor is received. Also, in step 2402, the transductive classifier is trained by iterative calculations using at least one predetermined cost factor, at least one seed image, and unlabeled images, where for each iteration of the calculation The cost factor is adjusted as a function of the expected label value. After at least some iterations, at step 2404, a confidence score for the unlabeled seed image is stored.

  Further, at step 2406, the identifier of the unlabeled document with the highest confidence score is output to at least one of the user, another system, and another process. The identifier can be an electronic copy of the document itself, their parts, their title, their name, their file name, a pointer to the document, and so on. Also, after each iteration, a confidence score may be stored, and an identifier of the unlabeled image having the highest confidence score is output after each iteration. In addition, prior probabilities of data point labels for labeled and unlabeled images can be received, and for each iteration of the calculation, the prior probabilities of data point labels can be adjusted according to an estimate of the class membership probability of the data points. . In addition, a third unlabeled image of the face can be received, eg, from the airport checkpoint example described above, and this third unlabeled image is compared to at least some images having the highest confidence score. If the reliability of the face of the third unlabeled image is the same as the face of the seed image, the identifier of the third unlabeled image can be output.

  Yet another embodiment of the invention allows users to improve their search results by providing feedback to the document discovery system. For example, when performing a search on a search engine on the Internet, such as a search result for patent documents or patent application documents, a user may obtain a number of results in response to his search query. One embodiment of the present invention re-examines the results suggested by the search engine from the user and the relevance of one or more retrieved results, eg, “close to but not the one I wanted. ”,“ Completely different ”, etc. can be reported to the engine. Each time a user provides feedback to the engine, better results are prioritized for user review.

  A document retrieval method in one embodiment is presented in FIG. At step 2500, a search query is received. The search query can be any type of query, including case sensitive queries, Boolean queries, approximate matching queries, structured queries, and so on. In step 2502, a document based on the search query is retrieved. In addition, a document is output in step 2504, and a user input label indicating the relevance of the document to the search query is received in step 2506 for at least some documents. For example, the user may indicate whether a particular result returned from the query is relevant. Also, at step 2508, the classifier is trained based on the search query and the user input label, and at step 2510, a document classification technique is performed on the document using the classifier to reclassify the document. Further, at step 2512, at least some document identifiers are output based on the document classification. The identifier can be an electronic copy of the document itself, their parts, their title, their name, their file name, a pointer to the document, and so on. The reclassified document can also be output along with the most reliable document that was output first.

  Document classification techniques may include any type of processing, such as transductive processing, support vector machine processing, maximum entropy identification processing, and the like. Any inductive or transductive approach described above can be used. In a preferred approach, the classifier can be a transductive classifier, and the transductive classifier can be trained by iterative computation using at least one predetermined cost factor, a search query, and a document, where For each iteration of the calculation, the cost factor can be adjusted as a function of the expected label value and a trained classifier can be used to classify the document. Further, the prior probabilities of the data point labels for the search query and the document may be received, and for each iteration of the calculation, the prior probabilities of the data point labels may be adjusted according to the estimate of the class membership probability of the data points.

  Further embodiments of the invention can be used to improve ICR / OCR and speech recognition. For example, many embodiments of speech recognition programs and speech recognition systems require an operator to repeat several words in order to train the system. The present invention may first collect user unclassified content by monitoring the user's voice for a preset period, for example by listening to telephone conversations. As a result, when the user starts training the recognition system, the system supports the construction of the memory model by using the sound monitored by utilizing transductive learning.

  In yet another embodiment, a method for verifying the association between an invoice and an entity is presented in FIG. At step 2600, the classifier is trained based on the type of invoice associated with the first entity. The invoice format may refer to either or both of the physical layout of the indicia on the invoice or features such as keywords, invoice number, customer name, etc. on the invoice. Further, at step 2602, a plurality of invoices labeled as associated with at least one of the first entity and the other entity are accessed, and at step 2604, a classifier is used to relate the invoice. A document classification technique is executed. For example, any inductive or transductive technique described above can be used as the document classification technique. For example, document classification techniques may include transductive processing, support vector machine processing, maximum entropy identification processing, and the like. Also, at step 2606, at least one identifier of the invoice having a high probability that is not associated with the first entity is output.

  Further, the classifier can be any type of classifier, eg, a transductive classifier, which uses at least one predetermined cost factor, at least one document classification, and an invoice, It can be trained by iterative calculations, and for each iteration of the calculation, the cost factor is adjusted as a function of the expected label value and the invoice is classified using a trained classifier. Also, data point label prior probabilities for the seed document and invoice may be received, and for each iteration of the calculation, the data point label prior probabilities are adjusted according to the data point class membership probability estimate.

  One advantage provided by the embodiments described herein is the stability of the transductive algorithm. This stability is achieved by scaling cost factors and adjusting the prior probabilities of labels. For example, in one embodiment, the transductive classifier is trained by iterative classification using at least one cost factor, labeled data points, and unlabeled data points as training examples. For each iteration of the calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value. Further, for each iteration of the calculation, the prior probabilities of the data point labels are adjusted according to the estimation of the data point class membership probability.

A workstation may have an operating system, such as a Microsoft Windows operating system (OS), MAC OS, or UNIX operating system. It will be appreciated that the preferred embodiments may also be implemented on platforms and operating systems other than those mentioned. Preferred embodiments may be described using JAVA®, XML, C, and / or C ⁺⁺ languages, or other programming languages, and also object-oriented programming methodologies. Object oriented programming (OOP), which is increasingly used to develop complex applications, can be used.

  The applications described above use transductive learning to overcome the problem of extremely sparse data that poses difficulties for inductive face recognition systems. This aspect of transductive learning is not limited to this application and can be used to solve other machine learning problems due to sparse data.

  Those skilled in the art can devise variations that are within the scope and spirit of the various embodiments of the invention disclosed herein. Further, the various features of the embodiments disclosed herein may be used alone or in various combinations with each other and are intended to be limited to the specific combinations described herein. It has not been done. Accordingly, the claims are not limited by the illustrated embodiment.

FIG. 1 is a drawing of a chart plotting expected labels as a function of classification score, obtained by incorporating MED discriminative learning applied to label induction. 2A-2H are plots of a series of plots showing the iteration of the decision function calculation obtained by transductive MED learning. FIGS. 3A-3H are plots of a series of plots showing the iterations of the decision function computation obtained by improved transductive MED learning according to one embodiment of the present invention. FIG. 4 shows a control flow diagram for classification of unlabeled data according to one embodiment of the present invention using scaled cost factors. FIG. 5 shows a control flow diagram for classification of unlabeled data according to one embodiment of the present invention using user-defined prior probability information. FIG. 6 shows a detailed control flow diagram for classification of unlabeled data according to one embodiment of the present invention using maximum entropy identification with scaled cost factors and prior probability information. FIG. 7 is a network diagram illustrating a network architecture in which various embodiments described herein may be implemented. FIG. 8 is a system diagram of a typical hardware environment associated with a user device. FIG. 9 shows a block diagram of an apparatus according to an embodiment of the present invention. FIG. 10 shows in a flowchart the classification process performed according to one embodiment. FIG. 11 shows in a flowchart the classification process performed according to one embodiment. FIG. 12 shows in a flowchart the classification process performed according to one embodiment. FIG. 13 shows in a flowchart the classification process performed according to one embodiment. FIG. 14 shows in a flowchart the classification process performed according to one embodiment. FIG. 15 shows in a flowchart the classification process performed according to one embodiment. FIG. 16 illustrates in a flowchart the classification process performed according to one embodiment. FIG. 17 shows in a flowchart the classification process performed according to one embodiment. FIG. 18 shows in a flowchart the classification process performed according to one embodiment. FIG. 19 illustrates in a flowchart the classification process performed according to one embodiment. FIG. 20 shows in a flowchart the classification process performed according to one embodiment. FIG. 21 shows in a flowchart the classification process performed according to one embodiment. FIG. 22 shows a control flow diagram illustrating the method of one embodiment of the present invention as applied to the first document separation system. FIG. 23 shows a control flow diagram illustrating the method of one embodiment of the present invention as applied to the second separation system. FIG. 24 shows in a flowchart the classification process performed according to one embodiment. FIG. 25 illustrates in a flowchart the classification process performed in accordance with one embodiment. FIG. 26 illustrates in a flowchart the classification process performed according to one embodiment. FIG. 27 illustrates in a flowchart the classification process performed according to one embodiment. FIG. 28 illustrates in a flowchart the classification process performed according to one embodiment. FIG. 29 illustrates in a flowchart the classification process performed according to one embodiment.

Claims (55)

A method of data classification in a computer-based system, comprising:
Receiving labeled data points, wherein each of the labeled data points is a training example for a data point for which the data point is to be included in a specified category or excluded from a specified category. Having at least one label indicating whether the training is for an example data point;
Receiving unlabeled data points;
Receiving at least one predetermined cost factor for the labeled and unlabeled data points;
Training a transductive classifier using maximum entropy discrimination (MED) by iterative computation using the at least one cost factor and the labeled and unlabeled data points as training examples, comprising: For each iteration, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point label is adjusted according to the estimate of the class membership probability of the data point;
Applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and an input data point;
And outputting the classified data classification points, or derivatives thereof, the user, another system, at least one and processes,
Including the method.   The method of claim 1, wherein the function is an absolute value of the expected label of a data point. Further comprising the steps of receiving priori probability information labeled and unlabeled data points A method according to claim 1.   The method of claim 3, wherein the transductive classifier learns using prior probability information of the labeled and unlabeled data. Before and Kira bell with data the unlabeled data using the training examples according to their expected label, a Gaussian prior distribution to the decision function parameters given training examples are training examples and exclusion included The method of claim 1, comprising the further step of using to determine a decision function with minimal KL divergence.   The method of claim 1, comprising the further step of determining a decision function having a minimum KL divergence using a polynomial prior for the decision function parameters.   The method of claim 1, wherein the iterative step of training a transductive classifier is repeated until convergence of data values is reached.   The method of claim 7, wherein convergence is reached when a change in a decision function of the transductive classifier falls below a predetermined threshold.   The method of claim 7, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold.   The method of claim 1, wherein the labels of the included training examples have a value of +1 and the excluded training examples of labels have a value of −1.   The method of claim 1, wherein the label of the included example is mapped to a first number and the label of the excluded example is mapped to a second number. Storing the labeled data points in a memory of a computer;
Storing the unlabeled data points in a memory of a computer;
Storing the input data points in a memory of a computer;
Storing at least one predetermined cost factor of the labeled and unlabeled data points in a memory of a computer;
The method of claim 1, further comprising: A method of data classification comprising providing computer-executable program code that is deployed and executed on a computer system, comprising:
The program code is
At least one indicating whether each labeled data point is a training example for a data point to be included in the specified category or a training point for a data point excluded from the specified category Instructions for accessing the labeled data point stored in the memory of the computer having one label;
Instructions for accessing unlabeled data points from the computer's memory;
Instructions for accessing at least one predetermined cost factor of the labeled and unlabeled data points from a computer memory;
Instructions for training a maximum entropy identification (MED) transductive classifier by iterative computation using the at least one stored cost factor and stored labeled data points and stored unlabeled data points. there, for each iteration of the computation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, prior probability of data points Ru is adjusted according to the estimated value of the class membership probability of a data point, the instruction And
Instructions for applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and an input data point;
The classification classification of data points, or a derivative thereof, a user, a command for outputting another system, at least one and processes,
Comprising
Method.   The method of claim 13, wherein the function is an absolute value of the expected label of a data point. Further comprising the steps of accessing the a priori probability information of labeled and unlabeled data points stored in the memory of a computer, The method according to claim 13.   The method of claim 15, wherein for each iteration, the prior probability information is adjusted according to an estimate of a class membership probability of a data point.   Utilizing the labeled and unlabeled data as learning examples according to their expected labels and having a minimum KL divergence for the pre-distribution of decision function parameters given the included training examples and excluded training examples The method of claim 13, further comprising instructions for determining a decision function.   The method of claim 13, wherein the iterative step of training a transductive classifier is repeated until convergence of data values is reached.   The method of claim 18, wherein convergence is reached when the change in the decision function of the transductive classification falls below a predetermined threshold.   The method of claim 18, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold.   The method of claim 13, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of −1.   The method of claim 13, wherein the label of the included example is mapped to a first number and the label of the excluded example is mapped to a second number. A data processing device, the device comprising:
(I) Whether each labeled data point is a training example for a data point that should be included in the specified category or a training example for a data point excluded from the specified category Storing at least one of the labeled data points having at least one label; (ii) unlabeled data points; and (iii) at least one predetermined cost factor for the labeled and unlabeled data points, With two memories
Using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples, a transductive maximum entropy identification (MED) is used for the transductive classifier. A transductive classifier training device for iterative teaching, wherein at each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the label of the data point is A training device that is adjusted according to an estimate of the class membership probability of the data points;
With
A classifier trained by the transductive classifier training device is used to classify at least one of the unlabeled data points, the labeled data points, and the input data points;
The classification classification of data points, or a derivative thereof, the user, another data processing apparatus, is output to at least one and processes,
apparatus.   24. The apparatus of claim 23, wherein the function is an absolute value of the expected label of a data point.   24. The apparatus of claim 23, wherein the memory also stores prior probability information for labeled and unlabeled data points.   26. The apparatus of claim 25, wherein in each iteration of the MED calculation, the prior probability information is adjusted according to an estimate of a data point class membership probability. Using the labeled and unlabeled data as learning examples according to their expected labels, have a minimum KL divergence for the prior distribution of decision function parameters given the included training examples and excluded training examples. further comprising a processor for determining a decision functions that, according to claim 23.   24. The apparatus of claim 23, further comprising means for determining convergence of the data value and terminating the calculation simultaneously with the determination of convergence.   29. The apparatus of claim 28, wherein convergence is reached when a change in a decision function of the transductive classifier calculation falls below a predetermined threshold.   29. The apparatus of claim 28, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold.   24. The apparatus of claim 23, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1.   24. The apparatus of claim 23, wherein the labels of the included examples are mapped to a first number and the labels of the excluded examples are mapped to a second number. A program storage medium body computer readable, the medium is substantively perform the method of data classification embodying the method one or more programs of instructions executable by the computer,
Each labeled data point indicates whether it is a training example for a data point to be included in a specified category or a training example for a data point excluded from a specified category, at least Receiving the labeled data point having one label;
Receiving unlabeled data points;
Receiving at least one predetermined cost factor of the labeled and unlabeled data points;
Training a transductive classifier by iterative maximum entropy identification (MED) calculation using the at least one stored cost factor and stored labeled data points and stored unlabeled data points as training examples. Where in each iteration of the MED calculation, the cost factor of the unlabeled data point is adjusted as a function of the expected label value, and the prior probability of the data point is adjusted according to the estimate of the class membership probability of the data point. Steps,
Applying the trained classifier to classify at least one of the unlabeled data point, the labeled data point, and an input data point;
And outputting the classified data classification points, or derivatives thereof, the user, another program storage medium, at least one and processes,
Including
Program storage medium . The program storage medium according to claim 33, wherein the function is an absolute value of the expected label of a data point. 34. The program storage medium of claim 33, wherein the method further comprises storing prior probability information for labeled and unlabeled data points in a computer memory. 36. The program storage medium of claim 35, wherein in each iteration of the MED calculation, the prior probability information is adjusted according to an estimate of a data point class membership probability. The method uses the labeled and unlabeled data as learning examples according to their expected labels, and uses a minimum for a pre-distribution of decision function parameters given the included training examples and excluded training examples. It determined that having a KL divergence further encompasses automatic answering step to determine the constant function program storage medium of claim 33. 34. The program storage medium of claim 33, wherein the iterative step of training a transductive classifier is repeated until convergence of data values is reached. 40. The program storage medium of claim 38, wherein convergence is reached when a change in the transductive classification decision function falls below a predetermined threshold. 39. The program storage medium according to claim 38, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold. 34. The program storage medium of claim 33, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1. 34. The program storage medium of claim 33, wherein the label of the included example is mapped to a first numeric value and the label of the excluded example is mapped to a second numeric value. A method for classifying unlabeled data in a computer-based system, comprising:
Receiving a labeled data point, whether the data point is a training example for a data point to be included in a specified category or a training example for a data point excluded from a specified category Each of the labeled data points has at least one label indicating:
Receiving labeled and unlabeled data points;
Receiving pre-label probability information for labeled and unlabeled data points;
Receiving at least one predetermined cost factor for the labeled and unlabeled data points;
Determining an expected label for each labeled and unlabeled data point according to the prior probability of the label of the data point;
Until the data value converges, and generating a scaled cost value for each unlabeled data points in proportion to the absolute value of the expected label following substeps, namely, the data points,
Use the labeled and unlabeled data as training examples according to their expected labels to minimize KL divergence for the prior probability distribution of the decision function parameters given the included training examples and excluded training examples determined by calculating the constant function that, the step of training a classifier,
Using the trained classifier to determine a classification score for the labeled and unlabeled data points;
Calibrating the output of the trained classifier to class membership probabilities;
Updating the prior probability of the label of the unlabeled data point according to the determined class membership probability;
Determining the probability distribution of the label and margin using maximum entropy identification (MED) using the updated prior probability of the label and the previously determined classification score;
Using the previously determined probability distribution of labels to calculate a new expected label;
Updating the expected label for each data point by interpolating the new expected label with the expected label from the previous iteration;
Repeating steps,
And outputting the classification of the input data points, or a derivative thereof, a user, another system, at least one and processes,
Including the method.   44. The method of claim 43, wherein convergence is reached when a change in the decision function falls below a predetermined threshold.   44. The method of claim 43, wherein convergence is reached when a change in the determined expected label value falls below a predetermined threshold.   44. The method of claim 43, wherein the labels of the included training examples have a value of +1 and the labels of the excluded training examples have a value of -1. Having a known confidence level relating to the label allocation, receiving a-out document at least with one label,
Receiving an unlabeled document;
Receiving at least one predetermined cost factor;
Training a transductive classifier by iterative computation using the at least one predetermined cost factor, the at least one labeled document, and the unlabeled document, for each iteration of the computation: The cost factor is adjusted as a function of the expected label value; and
Storing a confidence score for the unlabeled document after at least some of the iterations;
And outputting the identifier of the unlabeled documents, the user, the system, at least one and processes with the highest confidence score,
And the identifier of the unlabeled document includes a copy of at least a part of the unlabeled document, the file location ID of the unlabeled document, the name of the unlabeled document, or the title of the unlabeled document . A way to classify documents. 48. The method of claim 47, wherein the at least one labeled document has a list of keywords.   48. The method of claim 47, wherein a confidence score is stored after each iteration, and an identifier for the unlabeled document having the highest confidence score is output after each iteration. Further comprising the steps of receiving a prior probability of the label of the data points for the labeled and unlabeled documents, A method according to claim 47, for each iteration of the calculation, the label of the data points The prior probability of is adjusted according to an estimate of the class membership probability of the data points. Having a known confidence level, receiving a-out images with labels of the at least one face,
Receiving an unlabeled image;
Receiving at least one predetermined cost factor;
Training a transductive classifier by iteration using the at least one predetermined cost factor, the at least one labeled image, and the unlabeled image, for each iteration of the computation: The cost factor is adjusted as a function of the expected label value; and
After at least a portion of the repeat, and storing the confidence score for unlabeled picture image,
And outputting the identifier of the label without an image, the user, the system, at least one and processes with the highest confidence score,
And the identifier of the unlabeled document includes a copy of at least a part of the unlabeled document, the file location ID of the unlabeled document, the name of the unlabeled document, or the title of the unlabeled document . Face recognition method. 52. The method of claim 51 , wherein the at least one labeled image has a label that indicates whether the image is included in a specified category. 52. The method of claim 51 , wherein a confidence score is stored after each of the iterations and an identifier of the unlabeled image having the highest confidence score is output after each iteration. Further comprising the steps of receiving a prior probability of the label of the data points for the labeled and unlabeled images, A method according to claim 51, for each iteration of the calculation, the label of the data points The prior probability of is adjusted according to an estimate of the class membership probability of the data points. Receiving a third unlabeled image of a face; comparing the third unlabeled image with at least a portion of the image having the highest confidence score; and a face of the third unlabeled image of when the reliability is identical to the face of said labeled image, further comprising the step of outputting an identifier of the third unlabeled image the method of claim 51.

Images