WO2019155523A1 - Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme - Google Patents

Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme Download PDF

Info

Publication number
WO2019155523A1
WO2019155523A1 PCT/JP2018/003994 JP2018003994W WO2019155523A1 WO 2019155523 A1 WO2019155523 A1 WO 2019155523A1 JP 2018003994 W JP2018003994 W JP 2018003994W WO 2019155523 A1 WO2019155523 A1 WO 2019155523A1
Authority
WO
WIPO (PCT)
Prior art keywords
classifier
matrix
processing sequence
label
evaluation
Prior art date
Application number
PCT/JP2018/003994
Other languages
English (en)
Japanese (ja)
Inventor
圭吾 木村
Original Assignee
日本電気株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電気株式会社 filed Critical 日本電気株式会社
Priority to PCT/JP2018/003994 priority Critical patent/WO2019155523A1/fr
Publication of WO2019155523A1 publication Critical patent/WO2019155523A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N99/00—Subject matter not provided for in other groups of this subclass

Definitions

  • the present disclosure relates to a classifier forming apparatus, a classifier forming method, and a non-transitory computer readable medium storing a program.
  • Multi-label classification is a classification problem that allows one data to belong to a plurality of categories (labels) at the same time. That is, the classification problem of “multi-label classification” is a problem of estimating one or more labels attached to data that has not yet been labeled (that is, unclassified data). In other words, the classification problem of “multi-label classification” is a problem of forming a classifier that calculates one or more labels attached to unclassified data.
  • Label Space Dimension Reduction Label Space Dimension Reduction
  • This label space reduction method is a method for solving a multilevel problem with relatively high accuracy by utilizing the existence of an “appearance influence relationship” between labels.
  • the “appearance influence relationship between labels” is a relationship in which the probability that the first label is attached to the data affects the probability that the second label is attached (that is, Label-Dependency).
  • the label space reduction method decomposes “label matrix Y” into a product of “feature matrix X” and “low-dimensional matrix U, V” as shown in the following equation (1).
  • “*” indicates a matrix product.
  • the “label matrix Y” is a binary matrix of N rows and L columns in which whether or not N data has each label of L labels is expressed by zero or one.
  • the “label matrix Y” is a matrix of N rows and L columns, each row corresponding to one data, each column corresponding to one label, and each matrix element corresponding to each matrix element. It is a value indicating whether or not a label corresponding to each matrix element is given to the data. That is, the value of the element Y ij of the label matrix Y is “1” if the i-th data has the j-th label, and “0” otherwise.
  • the “feature matrix X” is a matrix of N rows and F columns representing the F-dimensional features of N pieces of data.
  • the “feature matrix X” is a matrix of N rows and F columns, and each row corresponds to one data and each column corresponds to one feature. That is, the value of the element X ij of the “feature matrix X” is the value of the j th feature of the i th data.
  • the multilevel problem can be solved with relatively high accuracy by using the “appearance influence relationship between labels” by “low rank approximation” by matrix decomposition.
  • the present inventor treats all labels and all data equally in the related technology, that is, it is equivalent to performing uniform weighting on all labels and all data. Therefore, it has been found that there is a possibility that the classification accuracy for a label with a low appearance frequency is low.
  • the present inventor increases the accuracy of approximation only for labels with a high appearance frequency during low rank approximation, so that a classifier with a high classification accuracy is constructed only for labels with a high appearance frequency. I found that there is a possibility. And this inventor discovered that the classification precision about the label with low appearance frequency may not satisfy a desired classification standard in the related technology.
  • An object of the present disclosure is to provide a classifier forming apparatus, a classifier forming method, and a non-transitory computer-readable medium storing a program, which can form a classifier that improves classification accuracy.
  • J is a natural number
  • a classifier (J) including a first matrix of F rows and K columns and a second matrix of K rows and L columns is formed, and the feature matrix is a matrix of N rows and F columns, and each row corresponds to one data.
  • Each column corresponds to one feature
  • the label matrix is an N-row L-column matrix
  • each row corresponds to the one data
  • each column corresponds to one label
  • each matrix element In the data corresponding to each matrix element, a value indicating whether or not a label corresponding to each matrix element is given
  • the classifier formation From the first classifier formation processing sequence
  • the classifier coupling portion forming a bond classifier (J)
  • each classifier in the classifier forming processing sequence Finished with respect to the weight matrix output unit for outputting the weight matrix W different in the forming process sequence to the classifier forming unit, the classifier forming unit, the classifier combining unit, and the weight matrix output unit
  • the classifier formation processing sequence is repeated until a condition is satisfied, and the combined classifier formed by the classifier combination unit in the classifier formation processing sequence when the end condition is satisfied is defined as a learned classifier.
  • the classifier forming method in the J (J is a natural number) classifier forming processing sequence, a feature matrix and an answer label matrix that are learning data, and a weight matrix W that is different in each classifier forming processing sequence.
  • a classifier (J) including a first matrix of F rows and K columns and a second matrix of K rows and L columns is formed, and the feature matrix is a matrix of N rows and F columns, and each row is 1 Corresponding to one data and each column corresponding to one feature, and said label matrix is a matrix of N rows and L columns, each row corresponds to said one data and each column corresponds to one label and
  • Each matrix element is a value indicating whether or not a label corresponding to each matrix element is attached to data corresponding to each matrix element, and in the J-th classifier formation processing sequence, the classifier forming unit
  • the first classifier formation process by By combining the J classifier formed by the J-th classifier forming process sequence from the scan, coupled classifier (J) is formed
  • the non-transitory computer-readable medium is different in each classifier formation processing sequence from a feature matrix and an answer label matrix as learning data in a J (J is a natural number) classifier formation processing sequence.
  • a classifier (J) including a first matrix of F rows and K columns and a second matrix of K rows and L columns is formed, and the feature matrix is a matrix of N rows and F columns, Each row corresponds to one data and each column corresponds to one feature, and the label matrix is a matrix of N rows and L columns, each row corresponds to the one data, and each column corresponds to one label
  • Each matrix element is a value indicating whether or not the data corresponding to each matrix element is given a label corresponding to each matrix element, and in the J-th classifier formation processing sequence, the classification 1st classification by vessel forming part
  • the forming processing sequence combining the J-number of classifiers formed by the J-th classifier forming processing sequence, to form a bond classifier (J), said classifier
  • the present disclosure can provide a classifier forming apparatus, a classifier forming method, and a non-transitory computer-readable medium storing a program, which can form a classifier that improves classification accuracy.
  • FIG. 6 is a schematic diagram for explaining the process of combining the first matrix U. It is a schematic diagram with which it uses for description of the joint process of 2nd matrix V. It is a flowchart which shows an example of the processing operation of the classifier formation apparatus of 1st Embodiment. It is a block diagram which shows an example of the classifier formation apparatus of 2nd Embodiment. It is a figure where it uses for description of calculation of the weighting matrix W using the right-wrong evaluation matrix Z. It is a flowchart which shows an example of the processing operation of the classifier formation apparatus of 2nd Embodiment. It is a figure which shows the hardware structural example of a classifier formation apparatus.
  • FIG. 1 is a block diagram illustrating an example of a classifier forming apparatus according to the first embodiment.
  • the classifier forming apparatus 10 illustrated in FIG. 1 repeatedly executes the “classifier adjustment processing sequence” until the “end condition” is satisfied.
  • the classifier forming apparatus 10 forms a “classifier” (that is, a weak classifier) in each classifier adjustment processing sequence.
  • the classifier forming apparatus 10 is formed in each classifier adjustment processing sequence by a classifier (that is, weak classifier) formed by each classifier adjustment processing sequence itself and a classifier adjustment processing sequence before that. Are combined to form a “joint classifier” (ie, a strong classifier).
  • the classifier forming apparatus 10 outputs the combined classifier formed by the classifier adjustment processing sequence when the termination condition is satisfied as a “learned classifier”.
  • the classifier forming apparatus 10 includes a classifier forming unit 11, a classifier combining unit 12, a weight matrix output unit 13, and a classifier formation control unit 14.
  • the classifier formation control unit 14 causes the classifier formation unit 11, the classifier combination unit 12, and the weight matrix output unit 13 to repeat the “classifier formation processing sequence” until the “end condition” is satisfied. Then, the classifier formation control unit 14 outputs the “joined classifier” formed by the classifier combination unit 12 in the classifier formation processing sequence when the “end condition” is satisfied as the “learned classifier”. To do.
  • the “end condition” may include a condition that the value of the number of repetitions J of the classifier formation processing sequence has reached a predetermined value of 2 or more. That is, as the “end condition”, the size of a matrix that is a “joint classifier” may be used.
  • the “end condition” may include a condition that the classification accuracy of the obtained “joined classifier” exceeds a reference. The “end condition” will be described in detail later.
  • the weight matrix output unit 13 outputs a “weight matrix” for inputting to the classifier forming unit 11 in each classifier forming processing sequence.
  • the “weight matrix” output from the weight matrix output unit 13 is a non-negative matrix and is different in each classifier formation processing sequence.
  • the weight matrix output unit 13 holds (stores) a plurality of weight matrices having different weights for a plurality of labels including a label having a high appearance frequency and a label having a low appearance frequency. Then, the weight matrix output unit 13 changes the weight matrix to be output for each classifier formation processing sequence and outputs the weight matrix. For example, in the first classifier formation processing sequence, the weight matrix output unit 13 sets an initial weight matrix W 0 as a weight matrix (weight of appearance frequency) for uniformly weighting all combinations of labels and data. (Equivalent to a weight matrix that puts weight on high labels). In this case, the “initial weight matrix W 0 ” has values of all matrix elements “1”, for example.
  • the weight matrix output unit 13 outputs a weight matrix that places weights on labels with a low appearance frequency.
  • the classifier which put weight on the label with high appearance frequency and the classifier put weight on the label with low appearance frequency can be formed in a balanced manner by a plurality of classifier formation processing sequences.
  • the weight matrix output unit 13 of the first embodiment generates a weight matrix using the generated random number, thereby changing the output weight matrix for each classifier formation processing sequence and outputting the weight matrix. May be. Even in this way, a plurality of classifier formation processing sequences can be formed in a well-balanced manner, a classifier placing weight on a label with a high appearance frequency and a classifier placing weight on a label with a low appearance frequency.
  • the classifier forming unit 11 receives “feature matrix X” and “answer label matrix Y” as “learning data”. Further, the classifier forming unit 11 inputs the “weight matrix W” output from the weight matrix output unit 13. As described above, the “weight matrix” output from the weight matrix output unit 13 differs in each classifier formation processing sequence.
  • the classifier forming unit 11 performs “feature matrix X” and “answer label matrix Y” and the input weight matrix W in the J (J is a natural number) classifier forming processing sequence (that is, the Jth time). Based on the above, a classifier (J) including a “first matrix U J ” of F rows and K columns and a “second matrix V J ” of K rows and L columns is formed.
  • the “answer label matrix Y” is a binary matrix of N rows and L columns in which whether or not N data has each label of L labels is expressed by zero or one.
  • the “label matrix Y” is a matrix of N rows and L columns, each row corresponding to one data, each column corresponding to one label, and each matrix element corresponding to each matrix element. It is a value indicating whether or not a label corresponding to each matrix element is given to the data. That is, the value of the element Y ij of the label matrix Y is “1” if the i-th data has the j-th label, and “0” otherwise.
  • the “feature matrix X” is an N-row F-column matrix representing F-dimensional features of N pieces of data.
  • the “feature matrix X” is a matrix of N rows and F columns, each row corresponding to one data, each column corresponding to one feature, and each matrix element corresponding to each matrix element. Is a value indicating whether or not there is a feature corresponding to each matrix element. That is, the value of the element X ij of the “feature matrix X” is the value of the j th feature of the i th data.
  • the type of data is not particularly limited, and may be, for example, image data or document data as long as the data can be expressed by a feature vector.
  • the classifier forming unit 11 performs the weighted low rank approximation using the following equation (2), thereby performing “first matrix U J ” of F rows and K columns and “second” of K rows and L columns.
  • a classifier J classifier (U J , V J )) including a matrix V J ”is formed. That is, the classifier forming unit 11 performs a classifier including a “first matrix U J ” of F rows and K columns and a “second matrix V J ” of K rows and L columns by matrix decomposition using Expression (2) ( J) .
  • K K is a natural number
  • K is the rank of “first matrix U J ” and “second matrix V J ” obtained by matrix decomposition.
  • “ ⁇ ” represents a product for each matrix element
  • “*” represents a matrix product.
  • the classifier combination unit 12 includes J classifiers formed by the classifier forming unit 11 from the first classifier forming process sequence to the Jth classifier forming process sequence in the Jth classifier forming process sequence.
  • J join classifier
  • U ′ J , V ′ J join classifier
  • FIG. 2 is a schematic diagram for explaining the joining process of the first matrix U.
  • the first matrix U 1 and the first matrix U 2 of F rows and 1 column formed by the first and second classifier processing sequences are combined.
  • an F-row / 2-column coupling matrix U ′ 2 is formed.
  • the second matrix V 1 and the second matrix V 2 of 1 row L columns formed by the first and second classifier processing sequences are obtained.
  • a 2 ⁇ L connection matrix V ′ 2 is formed.
  • FIG. 4 is a flowchart illustrating an example of a processing operation of the classifier forming apparatus according to the first embodiment.
  • the classifier forming unit 11 inputs “feature matrix X” and “answer label matrix Y” (step S101).
  • the classifier formation control unit 14 causes the weight matrix output unit 13 to output the weight matrix W J-1 to the classifier formation unit 11 in the J-th classifier formation processing sequence (step S103). . That is, in the first classifier formation processing sequence, the “initial weight matrix W 0 ” is output from the weight matrix output unit 13.
  • the classifier forming unit 11 performs F row K K based on the “feature matrix X” and “answer label matrix Y” and the input weight matrix W J ⁇ 1 in the J- th classifier forming processing sequence.
  • the classifier combination unit 12 includes the J number of classifiers formed from the first classifier formation processing sequence to the Jth classifier formation processing sequence by the classifier formation unit 11 in the Jth classifier formation processing sequence.
  • a combined classifier (J) joint classifier (U ′ J , V ′ J )) is formed.
  • the classifier (1) itself formed in the first classifier formation processing sequence is used as the combined classifier (1). Treat as.
  • the classifier formation control unit 14 determines whether or not the “end condition” is satisfied (step S106).
  • step S106 NO When the “end condition” is not satisfied (step S106 NO), the classifier formation control unit 14 increments the value of J (step S107). Then, the processing step returns to step S103. Thereby, the next classifier formation processing sequence is started.
  • the classifier formation control unit 14 acquires the combined classifier (J) combined by the classifier combining unit 12 in the current classifier formation processing sequence. , And output as a “learned classifier” (step S108).
  • the weight matrix output unit 13 in the classifier forming apparatus 10 outputs a “weight matrix” for input to the classifier forming unit 11 in each classifier forming processing sequence. Then, the classifier forming unit 11 performs “feature matrix X” and “answer label matrix Y” and the input weight matrix W in the J (J is a natural number) classifier forming processing sequence (that is, the Jth time). Based on the above, a classifier (J) including a “first matrix U J ” of F rows and K columns and a “second matrix V J ” of K rows and L columns is formed.
  • a classifier that places a heavy weight on labels with a high appearance frequency and a classifier that places a heavy weight on a label with a low appearance frequency are formed in a balanced manner by a plurality of classifier formation processing sequences. Can do.
  • the classifier combination unit 12 performs the process from the first classifier forming process sequence to the Jth classifier forming process sequence by the classifier forming unit 11 in the Jth classifier forming process sequence.
  • a combined classifier (J) joint classifier (U ′ J , V ′ J )
  • a combined classifier (strong classifier) is formed from a plurality of classifiers (weak classifiers) formed in a well-balanced manner with respect to labels with high appearance frequency and labels with low appearance frequency. Can do. Thereby, a classifier that improves the classification accuracy can be formed.
  • FIG. 5 is a block diagram illustrating an example of a classifier forming apparatus according to the second embodiment.
  • the classifier forming apparatus 20 includes a weight matrix output unit 21 and a classifier formation control unit 22 in addition to the classifier forming unit 11 and the classifier combination unit 12.
  • the weight matrix output unit 21 includes an index calculation unit 21A and a weight matrix calculation unit 21B.
  • the index calculation unit 21A calculates an evaluation target label matrix Y ′ J based on the combined classifier (J) formed by the classifier combination unit 12 and the feature matrix X in the J-th classifier formation processing sequence. To do.
  • the index calculation unit 21A calculates the evaluation target label matrix Y ′ J using, for example, the following equation (4). That is, the evaluation target label matrix Y ′ J is a classification result obtained by classifying the feature matrix X using the combined classifier (J) .
  • the index calculating unit 21A calculates a correct / incorrect evaluation index (J) of the calculated evaluation target label matrix Y ′ J based on the calculated evaluation target label matrix Y ′ J and the answer label matrix Y. For example, index calculation unit 21A, by replacing the long each matrix element of the calculated evaluation label matrix Y 'J first less than the threshold replaced by "1" if the first threshold value or more "zero", 2 A valuated evaluation target label matrix Y ′ J is calculated.
  • the index calculation unit 21A assign a "zero" to the binarized evaluation label matrix Y 'correct matrix elements in comparison with the answer label matrix Y in J, incorrect A correct / wrong evaluation matrix Z J in which “1” is assigned to the matrix element is calculated. That is, the calculation of the correctness evaluation matrix Z J, can be used Hamming loss.
  • correctness evaluation matrix Z J is a matrix that determines binding classifiers for each combination of the data and the label (J) classification results using is incorrect or was right. Namely, each matrix element of the correctness evaluation matrix Z J is for the combination of the data and the label corresponding to each matrix element, shows the correctness of the classification using the binding classifier (J).
  • the correctness evaluation matrix Z J of the first classifier formation processing sequence is There is a high possibility that a matrix element corresponding to a label with a low appearance frequency indicates a classification error.
  • the weight matrix calculation unit 21B uses the weight matrix W used in the J + 1 classifier formation processing sequence based on the correctness / incorrectness evaluation index (J) calculated by the index calculation unit 21A in the Jth classifier formation processing sequence. J is calculated. For example, the weight matrix calculation unit 21B replaces one of correctness evaluation matrix Z J calculated by the index calculation unit 21A, greater than zero and less than 1 to the "first value”. The weight matrix calculation unit 21B has a zero in the correctness evaluation matrix Z J calculated by the index calculation unit 21A, replaced with less than greater than zero and "first value", "second value".
  • the weight matrix calculation unit 21B can calculate each matrix element W ij of the weight matrix W J from each matrix element Z ij of the accuracy evaluation matrix Z J by using the following equation (5).
  • equation (5) ⁇ and ⁇ are non-negative real numbers, respectively.
  • each matrix element Z ij of the correct / incorrect evaluation matrix Z J is “zero” if the j-th label of the i-th data is correctly classified by the joint classifier (J) . “1”.
  • the value of each matrix element W ij of the weight matrix W J is “ ⁇ ” if the classification for the j-th label of the i-th data is correctly performed by the joint classifier (J) .
  • the weighting matrix W J gives a larger weight to the combination of data and label that is incorrectly classified by the combined classifier (J). It can be performed.
  • the weighting matrix W J is used in the J + 1-th classifier forming process sequence. For this reason, in the J + 1 classifier formation processing sequence, the combination of the data and the label, in which the classification by the combined classifier (J) formed in the Jth classifier formation processing sequence is incorrect, is given more weight.
  • a placed classifier (J + 1) can be formed.
  • FIG. 6 is a diagram for explaining the calculation of the weight matrix W using the accuracy evaluation matrix Z.
  • "zero" in the correctness evaluation matrix Z J is replaced by “0.1”.
  • the classifier formation control unit 22 sets “end condition” to the classifier formation unit 11, the classifier combination unit 12, and the weight matrix output unit 21. The “classifier formation processing sequence” is repeated until it is satisfied.
  • FIG. 7 is a flowchart illustrating an example of the processing operation of the classifier forming apparatus according to the second embodiment.
  • the processing from step S201 to step S205 and step S209 to step S211 in FIG. 7 is basically the same as the processing from step S101 to step S108 in FIG.
  • the index calculation unit 21A calculates an evaluation target label matrix Y ′ J based on the combined classifier (J) formed by the classifier combination unit 12 and the feature matrix X in the J-th classifier formation processing sequence. (Step S206).
  • the index calculation unit 21A calculates a correct / incorrect evaluation index (J) of the calculated evaluation target label matrix Y ′ J based on the calculated evaluation target label matrix Y ′ J and the answer label matrix Y (step S207). .
  • the index calculator 21A is index calculation unit 21A, if the less than 1 threshold replace each matrix element of the calculated evaluation label matrix Y 'J to "1" if the first threshold value or more By replacing with “zero”, the binarized evaluation target label matrix Y ′ J is calculated.
  • the index calculation unit 21A as the correctness evaluation index (J), assign a "zero" to the binarized evaluation label matrix Y 'correct matrix elements in comparison with the answer label matrix Y in J, incorrect A correct / wrong evaluation matrix Z J in which “1” is assigned to the matrix element is calculated.
  • the weight matrix calculation unit 21B uses the weights used in the J + 1th classifier formation processing sequence based on the correct / incorrect evaluation index (J) calculated by the index calculation unit 21A in the Jth classifier formation processing sequence.
  • a matrix W J is calculated (step S208). Specifically, the weight matrix calculation unit 21B replaces one of correctness evaluation matrix Z J calculated by the index calculation unit 21A, greater than zero and less than 1 to the "first value”.
  • the weight matrix calculation unit 21B has a zero in the correctness evaluation matrix Z J calculated by the index calculation unit 21A, replaced with less than greater than zero and "first value", "second value".
  • the classifier formation control unit 22 determines whether or not the “end condition” is satisfied (step S209).
  • the weight matrix W J is calculated in step S208 before the determination about the “end condition” in step S209.
  • the present invention is not limited to this. Calculation of the weight matrix W J in step S208, when the "end condition" is not determined to be satisfied in step S209, may be executed.
  • the index calculating unit 21A in the classifier forming apparatus 20 includes the combined classifier (J) formed by the classifier combining unit 12 in the J-th classifier forming processing sequence. Based on the feature matrix X, the evaluation target label matrix Y ′ J is calculated. Then, the index calculating unit 21A calculates a correct / incorrect evaluation index (J) of the calculated evaluation target label matrix Y ′ J based on the calculated evaluation target label matrix Y ′ J and the answer label matrix Y. Then, the weight matrix calculation unit 21B uses the weights used in the J + 1th classifier formation processing sequence based on the correctness / incorrectness evaluation index (J) calculated by the index calculation unit 21A in the Jth classifier formation processing sequence. A matrix W J is calculated.
  • the classification by the combined classifier (J) formed in the Jth classifier forming processing sequence is incorrect.
  • a classifier (J + 1) can be formed with more weight on the combination.
  • the classifier which put weight on the label with high appearance frequency and the classifier put weight on the label with low appearance frequency can be formed in a balanced manner by a plurality of classifier formation processing sequences.
  • the “end condition” can include a condition based on the classification accuracy of the “joint classifier”.
  • the “end condition” may include a condition that the sum of the element values of the accuracy evaluation matrix Z J is equal to or less than the second threshold value.
  • the “termination condition” may include a condition that the average value of the sum of the element value of each column of correctness evaluation matrix Z J is equal to or less than the third threshold value.
  • the present invention is not limited to this.
  • the classifier (J) may be obtained using the L1 distance as in the following equation (6).
  • a “non-negative constraint” in which each matrix element of the classifier (U J , V J ) is non-negative may be applied.
  • ⁇ 2> a method of calculating each matrix element W ij of the weight matrix W J from each matrix element Z ij of the correctness / incorrectness evaluation matrix Z J by using the above equation (5).
  • the present invention is not limited to this. That is, in the above equation (5), common ⁇ and ⁇ are used for all combinations of data and labels, but the present invention is not limited to this.
  • the values of ⁇ and ⁇ may be changed for each label, and the values of ⁇ and ⁇ may be changed for each data. That is, weighting may be performed by changing the weight between labels, or may be performed by changing the weight between data.
  • each matrix element W ij of the weight matrix W J may be calculated from each matrix element Z ij of the accuracy evaluation matrix Z J using the following equation (7).
  • ⁇ and ⁇ are non-negative L-dimensional weight coefficient vectors.
  • correctness evaluation matrix Z J as correctness evaluation index (J), but is not limited thereto. That is, in the second embodiment, correctness / incorrectness evaluation is performed for each combination of data and label, but the present invention is not limited to this.
  • correctness evaluation may be performed in units of labels, and correctness evaluation may be performed in units of data.
  • an F value can be used. That is, the index calculation unit 21A, when performing the correctness evaluated by label basis as correctness evaluation index (J), correctness evaluation vector (J to the F value of each column in the calculated evaluation target label matrix Y 'J and vector elements ) May be calculated.
  • the true / false evaluation vector (J) is an L-dimensional vector.
  • index calculation unit 21A when performing the correctness evaluated in data units, as correctness evaluation index (J), correctness evaluation vector (J to the F value for each row in the calculated evaluation target label matrix Y 'J and vector elements ) May be calculated.
  • the accuracy evaluation vector (J) is an N-dimensional vector.
  • the weight matrix W J is described as being calculated using the correctness / incorrectness evaluation matrix Z J.
  • the present invention is not limited to this. That is, in the second embodiment, one weight is calculated for each combination of data and label, but the present invention is not limited to this.
  • the weight may be calculated in units of labels, or the weight may be calculated in units of data.
  • the weight matrix W can be calculated using the following equation (8). That is, when the F value corresponding to a certain label is substituted into Z j in the equation (8), the values of all matrix elements corresponding to the label are calculated.
  • the F value indicates that the larger the F value is, the more correctly the classification is performed. Therefore, in Expression (8), the weight of each label is calculated by subtracting the F value corresponding to each label from 1 which is the maximum value of the F value.
  • FIG. 8 is a diagram illustrating a hardware configuration example of the classifier forming apparatus.
  • the classifier forming apparatus 100 includes a processor 101 and a memory 102.
  • the classifier forming unit 11, the classifier combining unit 12, the weight matrix output units 13 and 21, and the classifier formation control units 14 and 22 of the classifier forming apparatuses 10 and 20 are stored in the memory 102 by the processor 101. It may be realized by reading and executing the program.
  • the program may be stored using various types of non-transitory computer readable media and provided to the classifier forming apparatus 100.
  • the program may be supplied to the classifier forming apparatus 100 by various types of temporary computer-readable media.
  • a classifier forming unit that is a value indicating whether or not a label corresponding to the element is given; In the J-th classifier forming processing sequence, by combining the J classifiers formed from the first classifier forming processing sequence to the J-th classifier forming processing sequence by the classifier forming unit.
  • a classifier combination forming a combined classifier (J) ;
  • a weight matrix output unit for outputting the weight matrix W, which is different in each classifier formation processing sequence in each classifier formation processing sequence, to input to the classifier formation unit;
  • the classifier forming unit, the classifier combining unit, and the weight matrix output unit repeat the classifier forming processing sequence until an end condition is satisfied, and the classifier is formed when the end condition is satisfied.
  • a classifier formation controller that outputs the combined classifier formed by the classifier combiner in the processing sequence as a learned classifier;
  • a classifier forming apparatus comprising:
  • the weight matrix output unit includes: In the J-th classifier formation processing sequence, an evaluation target label matrix (J) is calculated based on the formed combined classifier (J) and the feature matrix, and the calculated evaluation target label matrix (J ) And the answer label matrix, an index calculation unit that calculates a correct / incorrect evaluation index (J) of the calculated evaluation target label matrix (J) ; In the J-th classifier formation processing sequence, a weight matrix calculation unit that calculates a weight matrix W J used in the J + 1-th classifier formation processing sequence based on the calculated accuracy evaluation index (J) ; Comprising In the first classifier formation processing sequence, the classifier forming unit forms the classifier (J) based on the feature matrix, the answer label matrix, and an initial weight matrix W 0, and the second and subsequent classes. In the classifier formation sequence, the classifier (J) is determined based on the feature matrix and the answer label matrix and the weight matrix W J-1 calculated in the J-1th classifier formation processing sequence.
  • the index calculation unit assigns 0 to a matrix element that is erroneously assigned as a correct / incorrect evaluation index (J) by assigning 0 to a correct matrix element compared to the answer label matrix in the calculated evaluation target label matrix (J) .
  • Calculate the assigned correct / incorrect evaluation matrix (J) The classifier forming apparatus according to appendix 2.
  • the termination condition includes a condition based on the accuracy evaluation matrix (J) , The classifier forming apparatus according to appendix 3 or 4.
  • the termination condition includes a condition that a sum of element values of the accuracy evaluation matrix (J) is equal to or less than a threshold value.
  • the classifier forming apparatus according to any one of appendices 3 to 5.
  • the termination condition includes a condition that an average value of the sum of element values of each column of the accuracy evaluation matrix (J) is equal to or less than a threshold value.
  • the classifier forming apparatus according to any one of appendices 3 to 5.
  • the index calculation unit calculates a correct / incorrect evaluation vector (J) having an F value for each column in the calculated evaluation label matrix (J) as a vector element as the correct / incorrect evaluation index (J) .
  • the classifier forming apparatus according to appendix 2.
  • the termination condition includes a condition that the value of J has reached a predetermined value of 2 or more.
  • the classifier forming apparatus according to any one of appendices 1 to 8.
  • the first of F rows and K columns is based on a feature matrix and an answer label matrix that are learning data, and a weight matrix W that is different in each classifier formation processing sequence.
  • the answer label matrix is a matrix of N rows and L columns, each row corresponding to the one data, each column corresponding to one label, and each matrix element corresponding to each matrix element.
  • J-th classifier formation processing sequence by combining J classifiers formed from the first classifier formation processing sequence to the J-th classifier formation processing sequence, a combined classifier (J ) The classifier formation processing sequence is repeated until an end condition is satisfied, and a combined classifier formed in the classifier formation processing sequence when the end condition is satisfied is output as a learned classifier.
  • Classifier formation method by combining J classifiers formed from the first classifier formation processing sequence to the J-th classifier formation processing sequence, a combined classifier (J ) The classifier formation processing sequence is repeated until an end condition is satisfied, and a combined classifier formed in the classifier formation processing sequence when the end condition is satisfied is output as a learned classifier.
  • an evaluation target label matrix (J) is calculated based on the formed combined classifier (J) and the feature matrix, and the calculated evaluation target label matrix (J ) and on the basis of said answer label matrix, correctness evaluated to calculate an index (J) of the calculated evaluation target label matrix (J),
  • a weight matrix W J used in the J + 1-th classifier formation processing sequence is calculated based on the calculated accuracy evaluation index (J) .
  • the feature matrix and the answer label matrix and the initial weight matrix W 0 or the J ⁇ 1-th classifier formation processing sequence are calculated. Forming the classifier (J) based on the weighted matrix W J ⁇ 1 , The classifier formation method according to appendix 10.
  • the termination condition includes a condition based on the accuracy evaluation matrix (J) , 14.
  • the termination condition includes a condition that a sum of element values of the accuracy evaluation matrix (J) is equal to or less than a threshold value. 15.
  • the classifier forming method according to any one of appendices 12 to 14.
  • the termination condition includes a condition that an average value of the sum of element values of each column of the accuracy evaluation matrix (J) is equal to or less than a threshold value. 15. The classifier forming method according to any one of appendices 12 to 14.
  • the termination condition includes a condition that the value of J has reached a predetermined value of 2 or more. 18.
  • the first of F rows and K columns is based on a feature matrix and an answer label matrix that are learning data, and a weight matrix W that is different in each classifier formation processing sequence.
  • the answer label matrix is a matrix of N rows and L columns, each row corresponding to the one data, each column corresponding to one label, and each matrix element corresponding to each matrix element.
  • J-th classifier formation processing sequence by combining J classifiers formed from the first classifier formation processing sequence to the J-th classifier formation processing sequence, a combined classifier (J ) The classifier formation processing sequence is repeated until an end condition is satisfied, and a combined classifier formed in the classifier formation processing sequence when the end condition is satisfied is output as a learned classifier.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un dispositif de formation de classificateur, un procédé de formation de classificateur, et un support lisible par ordinateur non transitoire permettant de stocker un programme, qui peut former un classificateur dans lequel la précision de classification est améliorée. Dans un dispositif de formation de classificateur (10), une unité de sortie de matrice pondérée (13) effectue une sortie de façon à fournir la « matrice de pondération » en entrée d'une unité de formation de classificateur (11) dans chaque séquence de processus de formation de classificateur. De plus, dans une J-ième (à savoir, la J-ième fois) séquence de processus de formation de classificateur (où J est un nombre naturel), l'unité de formation de classificateur (11) forme un classificateur (J) contenant « une première matrice UJ » de F lignes et K colonnes et « une deuxième matrice VJ » de K lignes et L colonnes en fonction « d'une matrice de caractéristiques X », « d'une matrice d'étiquettes de solution Y », et de la matrice pondérée d'entrée W.
PCT/JP2018/003994 2018-02-06 2018-02-06 Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme WO2019155523A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/JP2018/003994 WO2019155523A1 (fr) 2018-02-06 2018-02-06 Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2018/003994 WO2019155523A1 (fr) 2018-02-06 2018-02-06 Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme

Publications (1)

Publication Number Publication Date
WO2019155523A1 true WO2019155523A1 (fr) 2019-08-15

Family

ID=67548254

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2018/003994 WO2019155523A1 (fr) 2018-02-06 2018-02-06 Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme

Country Status (1)

Country Link
WO (1) WO2019155523A1 (fr)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112182215A (zh) * 2020-09-27 2021-01-05 中润普达(十堰)大数据中心有限公司 一种基于涕液特征信息语义认知系统及其使用方法
CN112190269A (zh) * 2020-12-04 2021-01-08 兰州大学 基于多源脑电数据融合的抑郁症辅助识别模型构建方法
JP2023044240A (ja) * 2021-09-17 2023-03-30 大連理工大学 音声感情予測方法及びシステム
WO2023249555A3 (fr) * 2022-06-21 2024-02-15 Lemon Inc. Traitement d'échantillon basé sur un mappage d'étiquettes

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007157130A (ja) * 2005-12-06 2007-06-21 Mitsubishi Electric Research Laboratories Inc コンピュータにより実施される強分類器を構築する方法

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007157130A (ja) * 2005-12-06 2007-06-21 Mitsubishi Electric Research Laboratories Inc コンピュータにより実施される強分類器を構築する方法

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112182215A (zh) * 2020-09-27 2021-01-05 中润普达(十堰)大数据中心有限公司 一种基于涕液特征信息语义认知系统及其使用方法
CN112190269A (zh) * 2020-12-04 2021-01-08 兰州大学 基于多源脑电数据融合的抑郁症辅助识别模型构建方法
CN112190269B (zh) * 2020-12-04 2024-03-12 兰州大学 基于多源脑电数据融合的抑郁症辅助识别模型构建方法
JP2023044240A (ja) * 2021-09-17 2023-03-30 大連理工大学 音声感情予測方法及びシステム
JP7271827B2 (ja) 2021-09-17 2023-05-12 大連理工大学 音声感情予測方法及びシステム
WO2023249555A3 (fr) * 2022-06-21 2024-02-15 Lemon Inc. Traitement d'échantillon basé sur un mappage d'étiquettes

Similar Documents

Publication Publication Date Title
WO2019155523A1 (fr) Dispositif de formation de classificateur, procédé de formation de classificateur, et support lisible par ordinateur non transitoire permettant de stocker un programme
US9002101B2 (en) Recognition device, recognition method, and computer program product
Breuel Benchmarking of LSTM networks
CN112633311A (zh) 利用输入数据结构的高效黑盒对抗性攻击
JPH07296117A (ja) 減少された要素特徴部分集合を用いたパターン認識システム用の分類重みマトリックスを構成する方法
JPH06176202A (ja) 文字認識用の管理されたトレーニング増加多項式法および装置
Lambrou et al. Reliable probability estimates based on support vector machines for large multiclass datasets
US10346742B2 (en) Calculation device, calculation method, and recording medium
CN112633309A (zh) 基于贝叶斯优化的查询高效的黑盒对抗攻击
JP7028322B2 (ja) 情報処理装置、情報処理方法及び情報処理プログラム
Monteiro et al. Generalizable adversarial examples detection based on bi-model decision mismatch
US9536206B2 (en) Method and apparatus for improving resilience in customized program learning network computational environments
KR20220038907A (ko) 생성적 대립 신경망(gan) 기반의 데이터 예측 방법, 그리고 이를 구현하기 위한 장치
CN110110845B (zh) 一种基于并行多级宽度神经网络的学习方法
EP3683736A1 (fr) Procédé, programme et appareil d'apprentissage machine
JP2018106618A (ja) 画像データ分類装置、オブジェクト検出装置及びこれらのプログラム
CN112149825A (zh) 神经网络模型的训练方法及装置、电子设备、存储介质
CN110506280B (zh) 神经网络训练系统、方法和计算机可读存储介质
JP7502370B2 (ja) カスタマイズされたディープラーニングベースのテキスト修正のための方法及び装置
JP2021051589A5 (fr)
Anumasa et al. Improving robustness and uncertainty modelling in neural ordinary differential equations
US11526693B1 (en) Sequential ensemble model training for open sets
CN111783936A (zh) 卷积神经网络构建方法、装置、设备及介质
Mendonça et al. Adversarial training with informed data selection
US11636286B1 (en) Concurrent ensemble model training for open sets

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18905371

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18905371

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: JP