JP2016091374A - Parameter estimation device, category assignment device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately assign a category to data.SOLUTION: A latent-vector estimation unit 32 estimates a latent vector xof a word included in a multiset of each document on the basis of a parameter aindicating a degree to which each document in a document set Dcontributes to category assignment, a category tassigned to each document, and the document set Dsuch that the latent vector xand the parameter aoptimize an objective function indicating a degree of effectiveness for document categorization. A discriminant-function estimation unit 34 repeats estimation of the parameter aoptimizing the objective function on the basis of the latent vector x, the category tassigned to each document, and the document set Duntil a predetermined termination condition is satisfied.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a parameter estimation device, a category assignment device, a method, and a program, and more particularly, to a parameter estimation device, a category assignment device, a method, and a program for assigning a category to data.

  In recent years, a large amount of document data has been accumulated. For example, papers, newspaper articles, and web pages. Since it is difficult for a user to read all the documents and classify the documents, there is a need for an apparatus that automatically classifies these documents into a desired category. By developing this apparatus, not only can documents be efficiently classified and organized, but also documents required by the user can be efficiently searched.

  In many cases, a document is represented as a collection of words. Conventional document classification methods classify documents based on the similarity between sets of words (for example, Non-Patent Document 1). However, because words with different word faces but similar meanings (for example, “test” and “test”) are not reflected in the similarity, the document may not be correctly classified.

  As a method for solving such a problem, a method is known in which words having different word faces but similar meanings are reflected in similarity by introducing word latent vectors. As a method for acquiring a latent vector representation of a word, for example, a method using singular value decomposition (SVD) (for example, Non-Patent Document 2) and a method using word2vec (for example, Non-Patent Document 3) are known.

Japanese Patent Application Laid-Open No. 2013-246795

Thorsten Joachims, “Text categorization with Support Vector Machines: Learning with many relevant features,” Lecture Notes in Computer Science Volume 1398, pp137-142, 1998. Zhu, Jun, Amr Ahmed, and Eric P. Xing. “MedLDA: maximum margin supervised topic models for regression and classication.” Proceedings of the 26th Annual International Conference on Machine Learning. ACM, 2009. Muandet, Krikamol, et al. “Learning from distributions via support measure machines.” Advances in Neural Information Processing Systems. 2012.

  However, the methods of Non-Patent Document 2 and Non-Patent Document 3 do not contribute much to the document classification accuracy because the word latent vectors are calculated by preprocessing regardless of the identification.

  The present invention has been made to solve the above-described problems, and provides a parameter estimation apparatus, method, and program capable of estimating parameters of a discrimination function for accurately assigning categories to data. The purpose is to provide.

  It is another object of the present invention to provide a category assignment apparatus, method, and program that can assign categories to data with high accuracy.

  In order to achieve the above object, a parameter estimation device according to a first aspect of the present invention is an input data set in which categories are assigned in advance, and data that can be expressed by multiple sets of discrete features. A latent vector of each discrete feature included in the multiple set of each data of the data set, a parameter indicating a degree of contribution of each data of the data set to category assignment, and a category assigned to each of the data of the data set And a latent vector of each discrete feature used in an identification function for identifying a category to be assigned to the data using the data set, and a parameter representing the degree to which each of the data of the data set contributes to category assignment An initialization unit that initializes and the degree of contribution of each of the data of the data set to category assignment A parameter representing the degree of contribution of each of the data of the discrete features and the latent vector of each discrete feature and the data of the data set based on the parameters, the category assigned to each of the data of the data set, and the data set , A latent vector estimator for estimating a latent vector of each discrete feature, a latent vector of each discrete feature, and a data of the data set of the data set so as to optimize an objective function representing an effective degree for categorizing the data. A discriminant function estimator for estimating a parameter representing a degree of contribution of each data of the data set to category assignment so as to optimize the objective function based on the category assigned to each and the data set And until the predetermined repetition termination condition is satisfied, the estimation by the latent vector estimation unit, and the discrimination function A repetition determining unit repeating estimation by tough, is configured to include a.

In the parameter estimation device according to the first invention, the discriminant function is represented by f (w _k ^* ) shown in the following equation, and the objective function is represented by L (A, X, It may be expressed by γ).

Where w represents the multiple set of data, w ^* represents the multiple set of data to which no category is assigned, and A represents a parameter a _i representing the degree to which each piece of data in the data set contributes to category assignment. Represents a set, 0 ≦ a _i ≦ C, t _i represents a category pre-assigned to the data, X represents a set of latent vectors x _v of each discrete feature v,

Represents a q-dimensional latent vector of the discrete feature v, γ is a parameter for controlling the similarity value between the discrete features, and K (w _i , w _j ; X, γ) is data i and data This is a kernel function that represents the degree of similarity with j.

  A category allocating device according to a second aspect of the present invention is an input unit that receives an input unknown data set that is a data set that can be expressed by multiple sets of discrete features and whose assigned category is unknown, and A latent vector of each discrete feature estimated by the parameter estimation device according to the first aspect of the invention, a parameter representing a degree of contribution of each of the data of the data set to category assignment, and each of the data of the data set An allocating unit that allocates the category according to the identification function to each data of the unknown data set received by the input unit based on the category and the data set.

  The parameter estimation method according to a first aspect of the present invention is the data set of data that is input by the initialization unit and to which categories are assigned in advance, and that can be expressed by multiple sets of discrete features. A latent vector of each discrete feature included in the multiple set of each of the data, a parameter representing a degree to which each of the data of the data set contributes to category assignment, a category assigned to each of the data of the data set, and the Using a data set, initial values of a latent vector of each discrete feature used in an identification function for identifying a category to be assigned to the data and a parameter indicating a degree of contribution of each data of the data set to category assignment And a latent vector estimator that determines how much each of the data in the data set contributes to category assignment. , A category assigned to each of the data of the data set, and a latent vector of each discrete feature and a degree to which each of the data of the data set contributes to category assignment based on the data set. Estimating a latent vector of each discrete feature such that a parameter to represent optimizes an objective function representing a degree of effectiveness in categorizing data; and a discriminant function estimator includes the latent vector of each discrete feature; Based on the category assigned to each data of the data set and the data set, a parameter representing the degree to which each of the data of the data set contributes to the category assignment is estimated so as to optimize the objective function. And the latent vector estimation unit until the repetition determination unit satisfies a predetermined repetition termination condition. According estimation, and wherein the performing comprises the steps of: repeating the estimation by the identification function estimator.

In the parameter estimation method according to the first invention, the discriminant function is represented by f (w _k ^* ) shown in the following equation, and the objective function is expressed as L (A, X, It may be expressed by γ).

Where w represents the multiple set of data, w ^* represents the multiple set of data to which no category is assigned, and A represents a parameter a _i representing the degree to which each piece of data in the data set contributes to category assignment. Represents a set, 0 ≦ a _i ≦ C, t _i represents a category pre-assigned to the data, X represents a set of latent vectors x _v of each discrete feature v,

Represents a latent vector of the q-dimensional discrete feature v, γ is a parameter for controlling the similarity value between the discrete features, and K (w _i , w _j ; X, γ) is data i and data This is a kernel function that represents the degree of similarity with j.

  In the category assigning method according to the second invention, the input unit accepts an input unknown data set that is a data set of data that can be expressed by multiple sets of discrete features and whose assigned category is unknown. A parameter representing a degree of contribution of each of the data of the discrete features estimated by the parameter estimation device according to claim 1 or 2 and the data of the data set to category allocation; Assigning the category according to the identification function to each data of the unknown data set received by the input unit based on the category assigned to each of the data of the set and the data set. It is characterized by executing in.

  A program according to a third invention is a program for causing a computer to function as each part of the parameter estimation device according to the first invention or the category assignment device according to the second invention.

  According to the parameter estimation apparatus, method, and program of the present invention, based on a parameter that represents the degree to which each of the data of the data set contributes to category assignment, the category assigned to each of the data, and the data set, A latent vector of each discrete feature included in each data multiplex set, and a parameter representing the degree to which each piece of data contributes to category assignment, optimizes an objective function that represents the degree to which the data is categorized effectively. Estimating a latent vector and estimating a parameter representing a degree of contribution to the category assignment so as to optimize the objective function based on the latent vector, the category assigned to each of the data, and the data set. By repeating until a predetermined repetition end condition is satisfied, the information for assigning the category accurately can be obtained. It is possible to estimate the parameters of the function, the effect is obtained that.

  In addition, according to the category assigning device, method, and program, each of the latent vector of each discrete feature included in the multiple sets of each data of the data set estimated by the parameter estimating device and the data of the data set are assigned to the category assignment. Based on a parameter representing the degree of contribution, a category assigned to each of the data, and a data set, a category is assigned to each data of the unknown data set according to an identification function. The effect that the categories can be assigned with high accuracy is obtained.

It is a block diagram which shows the functional structure of the parameter estimation apparatus and category allocation apparatus which concern on this Embodiment. It is a flowchart which shows the parameter estimation processing routine in the parameter estimation apparatus which concerns on this Embodiment. It is a flowchart which shows the category allocation process routine in the category allocation apparatus which concerns on this Embodiment. It is a graph which shows an example of the experimental result which allocated the category to the document using the method which concerns on this Embodiment. It is an abstract figure which shows an example of the visualization in the experimental result which allocated the category to the document using the method which concerns on this Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Configuration of Parameter Estimation Device According to Embodiment of the Present Invention>

  Next, the configuration of the parameter estimation apparatus according to the embodiment of the present invention will be described. In the present embodiment, a case where the present invention is applied to a parameter estimation apparatus that estimates a parameter for assigning a category to a document will be described as an example. However, as in the case of a document, input data is multiplexed with discrete features. The present invention can be applied to any data that can be expressed as a set. For example, the present invention can be applied to images and music.

  As shown in FIG. 1, a parameter estimation apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing a parameter estimation processing routine described later and various data. Can be configured with a computer. Functionally, the parameter estimation apparatus 100 includes an input unit 10, a calculation unit 20, and a parameter set storage unit 40 as shown in FIG.

  The input unit 10 is a set of documents in which categories are assigned in advance as input data.

Accept. Here, w _i is a multiple set of discrete features representing the i-th document (i = 1,..., N _tr ). For example, a word can be used as the discrete feature. In this case, w _i is a subset of the vocabulary set V. Also, t _i ε {+ 1, −1} represents the binary category label of the document i. Even when the category label is multi-valued, it can be dealt with by a method called a one-to-one method or a one-to-other method.

  The calculation unit 20 includes an initialization unit 30, a latent vector estimation unit 32, a discrimination function estimation unit 34, and an iterative determination unit 36.

The initialization unit 30 includes the latent vector x _v (vεV) of each word v included in the multiple set of words of each document of the document set D _tr received by the input unit 10 and each document of the document set D _tr parameters _a i indicating the degree to contribute to the category assigned _{(i = 1, ···, N} tr), document set _{D tr} category _{t i} assigned to each document, and using a document set _{D tr,} document A parameter a representing the degree of contribution of each of the latent vectors x _v of each word _v and the documents in the document set D _tr used in the identification function of the following equation (1) for identifying the category to be assigned to _i and the parameter γ are initialized.

Here, w represents a multiple set of words representing a document, and w ^* represents a multiple set of words representing a document to which no category is assigned. X = {x _v } _v∈V represents a latent vector set of each word v,

Represents a q-dimensional latent vector of the word v. K (w _i , w _j ; X, γ) is a kernel function that represents the degree of similarity between the document i and the document j, taking the set X of the latent vectors x _v as an input.

In the present embodiment, the initialization unit 30 initializes the latent vector x _v and the parameter γ of each word v at random. In addition, a parameter a _i indicating the degree to which each document in the document set D _tr contributes to category assignment is set as C. C is a predetermined parameter that determines whether misidentification is allowed.

The latent vector estimation unit 32 includes a parameter a _i indicating the degree to which each document of the document set D _tr initialized by the initialization unit 30 or estimated by the discrimination function estimation unit 34 contributes to category assignment, and a document set Based on the category t _i assigned to each document of D _tr and the document set D _tr , the latent vector x _v of each word v and the degree to which each document of the document set D _tr contributes to the category assignment. The latent vector x _v and the parameter γ of each word v are estimated so that the parameter a _i to be expressed optimizes the objective function of the following equation (2) that represents the degree of effectiveness in document category classification.

  here,

Is a set of parameters a _i representing the degree to which each of the documents in the document set D _tr contribute to category assignment, the domain of a _i is 0 ≦ a _i ≦ C. Further, when a _i = 0, it means that the document i does not contribute to category assignment. ρ> 0 is a predetermined parameter. As the kernel function, for example, the following equation (3) can be used.

Here, γ> 0 is a parameter for controlling the similarity value between words. In addition to the kernel function of the above formula (3), a kernel function having a semi-definite property such as a linear kernel or a polynomial kernel can be used. The estimation in the latent vector estimation unit 32 is performed by updating the latent vector and the parameter so that the kernel between documents belonging to the same category has a large value, and otherwise becomes a small value. This is possible by minimizing the objective function of equation (2) above given a set A of parameters a _i . For example, by using an optimization method such as the quasi-Newton method, the latent vector _xv and the parameter γ of each word can be estimated. The value of the objective function, when the potential vector x _v of each word v is given, for multiple set w _i of the word of all the documents that are included in the category labeled document set D _tr, as long rely on category t _i correctly It becomes the optimum value.

The discriminant function estimation unit 34 includes the latent vector x _v of each word v estimated by the latent vector estimation unit 32, the parameter γ, the category t _i assigned to each document in the document set D _tr , and the document set D. Based on _tr , a parameter a _i representing the degree to which each document in the document set D _tr contributes to category assignment is estimated so as to optimize the objective function.

The discriminant function estimation unit 34 intuitively obtains a set A of parameters a _i in the above equation (2) so that the category t _i and the sign of the discriminant function in the above equation (1) match for each document i. .

This is in terms of a set X and a parameter γ of latent vectors x _v by estimation of the potential vector estimator 32 is given, it is possible by maximizing the objective function of equation (2). For example, the parameters of the discriminant function can be estimated by using an optimization method such as quadratic programming or SMO (Sequential Minimal Optimization).

The repetition determination unit 36 repeats the estimation by the latent vector estimation unit 32 and the estimation by the discrimination function estimation unit 34 until a predetermined repetition end condition is satisfied. As a specific repetition end condition, the number of repetitions or the magnitude of change of the objective function can be used, and the update is repeated until the value of the objective function converges. If it is determined that the objective function has converged, each of the set X and the parameter γ of the latent vector x _v estimated by the latent vector estimation unit 32 and each of the documents of the document set D _tr estimated by the discrimination function estimation unit 34 Is stored in the parameter set storage unit 40. The parameter set storage unit 40 stores the set A of parameters a _i representing the degree of contribution to category assignment and the document set D _tr received by the input unit 10.

<Configuration of Category Assignment Device According to Embodiment of the Present Invention>

  Next, the configuration of the category assignment device according to the embodiment of the present invention will be described. As shown in FIG. 2, a category assignment device 300 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing a category assignment routine described later and various data. Can be configured. Functionally, the category assignment apparatus 300 includes an input unit 310, a calculation unit 320, a parameter set storage unit 40, and an output unit 350 as shown in FIG.

  The input unit 310 uses an unknown document set whose assigned category is unknown as input data.

Accept.

  The calculation unit 320 includes a reading unit 330 and an allocation unit 332.

The reading unit 330 includes a parameter a representing the degree to which each of the set X of the latent vector x _v of each word v stored in the parameter set storage unit 40, the parameter γ, and the document of the document set D _tr contributes to category assignment. The set A of _i and the document set D _tr are read.

The assignment unit 332 reads a set of parameters a _i representing the degree to which each of the documents X of the latent vector x _v of each word v read by the reading unit 330, the parameter γ, and the document set D _tr contributes to category assignment. and a, the category assigned to each of the documents in the document collection D _tr, based on the document set D _tr, for each document of unknown document set D _te, according to the identification function of the equation (1), the category Assign. In this embodiment, the value of the discriminant function is calculated using the above equation (1) for multiple sets w _k ^* (k = 1,..., N _te ) of unknown document words to which no category is assigned. To do. Here, when the value of the discriminant function is f (w _k ^* ) ≧ 0, the category y _k of the multiplex set w _k ^* of words of the unknown document is +1, and otherwise −1.

Then, the assigning unit 332 calculates the category label set obtained by calculating for each of the multiple sets w _k ^* of words of the unknown document.

Is output by the output unit 350.

<Operation of Parameter Estimation Device According to Embodiment of the Present Invention>

Next, the operation of the parameter estimation apparatus 100 according to the embodiment of the present invention will be described. When the input unit 10 receives the document set D _tr to which the category is assigned, the parameter estimation apparatus 100 executes a parameter estimation processing routine shown in FIG.

First, in step S100, the document set D _tr received by the input unit 10 is acquired.

Next, in step S102, latent vectors x _v for each word v included in the multiplexed set of words of each document in the document set D _tr acquired in step _S100, each of the documents in the document collection D _tr contribute to category assignment A parameter a _i representing the degree and a parameter γ are initialized.

At step S104, it is initialized at step S102, or the parameters _{a i} of each of the documents of the estimated document set _{D tr} represents the degree to contribute to the category assigned in step S106 to be described later, each of the documents in the document collection _{D tr} Based on the category t _i assigned to and the document set D _tr , the latent vector x _v of each word v and the parameter a _i representing the degree to which each of the documents in the document set D _tr contributes to category assignment are: The latent vector x _v and the parameter γ of each word v are estimated so as to optimize the objective function of the above equation (2) representing the degree of effectiveness in document category classification.

In step S106, the latent vectors _{x v} for each word v which is estimated in step S104, a parameter gamma, and category _{t i} assigned to each of the documents in the document collection _{D tr,} based on the document set _{D tr,} A parameter a _i representing the degree to which each document in the document set D _tr contributes to category assignment is estimated so as to optimize the objective function.

In step S108, it is determined whether the objective function has converged. If it is determined that the objective function has not converged, the processes from step S104 to step S106 are repeated. If it is determined that the objective function has converged, the estimation is performed in step S104. have been latent vectors x _v document set X and parameter gamma, each of the document of the estimated document set D _tr in step S106 obtained in the set a, and the step S100 of the parameters a _i indicating the degree to contribute to the category allocation of The set D _tr is stored in the parameter set storage unit 40, and the process ends.

<Operation of Category Assignment Device According to Embodiment of the Present Invention>

Next, the operation of the category assignment device 300 according to the embodiment of the present invention will be described. When the input unit 310 receives the unknown document set _Dte , the category assignment device 300 executes a category assignment processing routine shown in FIG.

First, in step S300, the unknown document set _Dte received by the input unit 310 is acquired.

Next, in step S302, a parameter representing the degree of contribution of each of the set X, parameter γ, and document set D _tr of the latent vector x _v of each word v that is stored in the parameter set storage unit 40 to the category assignment. The set A of a _i and the document set D _tr are read.

In step S304, a set A of parameters a _i representing the degree to which each of the set X of latent vectors x _v of each word v, parameter γ, and documents in the document set D _tr read in step S304 contributes to category assignment; a category assigned to each of the documents in the document collection D _tr, based on the document set D _tr, for each document of unknown document set D _te, according to the identification function of the equation (1), assign a category .

  In step S306, the category assignment result in step S304 is output by the output unit 350, and the process ends.

<Experimental result>

  Next, experimental results performed using the method according to the present embodiment will be described.

  In order to evaluate the method according to the present embodiment, an experiment was performed using three data sets (WebKB, Reuters-21578, and 20 Newsgroups). As comparison methods, MedLDA (Maximum entropy discrimination Latent Dirichlet Allocation) (see Non-Patent Document 2 above), SMM (Support Measure Machine) (above Non-Patent Document) after obtaining a vector representation of a word by singular value decomposition (SVD) Vector representation of words by SVD + SMM, word2vec (see Non-Patent Document 4 (Mikolov, Tomas, et al. “Distributed representations of words and phrases and their compositionality.” Advances in Neural Information Processing Systems. 2013.)) SVM (Support Vector Machine) using word2vec + SMM and rbf kernel and quadratic polynomial (poly2) kernel (Cortes, Corinna, and Vladimir Vapnik. “Support-vector networks . ”Machine learning 20.3 (1995): 273-297.)).

  FIG. 4 shows the classification accuracy when the number of documents to which categories are assigned is changed in learning by the method according to the present embodiment and learning by the comparison method described above. In all three data sets, the method according to the present embodiment (Lentent SMM) achieves the highest accuracy, suggesting that document classification is possible with high accuracy.

  In addition, FIG. 5 shows an example of word visualization results obtained by the method according to the present embodiment. Since the words of the same category are visualized together, this suggests that characteristic words can be efficiently found in each category.

As described above, according to the parameter estimation device according to the present embodiment, the parameter a _i indicating the degree to which each document in the document set D _tr contributes to category assignment, and the category assigned to each document. Based on t _i and the document set D _tr , the word latent vector x _v and the parameter a _i included in the multiple sets of each document optimize the objective function representing the degree of effectiveness in document category classification. as estimates potential vector x _v, and latent vectors x _v, and category t _i assigned to each document, on the basis of the document set D _tr, so as to optimize the objective function, the category assignment to estimate the parameters a _i representing the degree of contribution to, by repeating until the predetermined repeat end condition is satisfied, the identification for assigning accurately category It is possible to estimate the parameters of the number.

Further, according to the category assignment device according to the present embodiment, each word latent vector x _v included in the multiple set of each document in the document set estimated by the parameter estimation device and each document in the document set are classified into categories. Based on the parameter a _i indicating the degree of contribution to the assignment, the category t _i assigned to each of the documents, and the document set D _tr , for each document in the unknown document set D _te , according to the discriminant function: By assigning categories, categories can be assigned to documents with high accuracy.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the above embodiment, the description has been made assuming a document as data, but the present invention is not limited to this. If data that can represent one piece of data as a multiple set of discrete features is used, an image, Needless to say, it can also be applied to data such as music. Examples of the local feature of the image include SIFT (Scale Invariant Feature Transform) calculated from gradient information. Since a plurality of SIFTs are extracted from one image, a set whose elements are those discretized using vector quantization is a multiple set of discrete features of the image. If this multiple set is applied to the discriminant function of the above equation (1) as w, the parameters of the function for identifying the category from the set of images to which the category is assigned can be estimated, A category can be assigned to a set of images to which no is assigned.

10, 310 Input unit 20, 320 Operation unit 30 Initialization unit 32 Latent vector estimation unit 32 Latent vector estimation unit 34 Discrimination function estimation unit 34 Discrimination function estimation unit 36 Repetition determination unit 40 Parameter set storage unit 100 Parameter estimation device 300 Category allocation device 330 reading unit 332 allocation unit 350 output unit

Claims (7)

A latent vector of each discrete feature included in the multiple set of each data of the data set, which is a data set to which categories are assigned in advance and that can be represented by multiple sets of discrete features. A parameter representing a degree of contribution of each data of the data set to category assignment, a category assigned to each of the data of the data set, and a category assigned to the data using the data set An initialization unit for initializing a latent vector of each discrete feature and a parameter representing a degree that each of the data of the data set contributes to category assignment, which is used in an identification function;
A latent vector of each discrete feature based on a parameter representing the degree to which each of the data of the data set contributes to category assignment, a category assigned to each of the data of the data set, and the data set; and A latent vector estimator that estimates the latent vector of each discrete feature so that the parameter representing the degree to which each of the data in the data set contributes to category assignment optimizes the objective function representing the degree to which the data is categorized effectively When,
Each of the data in the data set is assigned a category so as to optimize the objective function based on the latent vector of each discrete feature, the category assigned to each of the data in the data set, and the data set. A discriminant function estimator for estimating a parameter representing the degree of contribution to
A repeat determination unit that repeats the estimation by the latent vector estimation unit and the estimation by the discrimination function estimation unit until a predetermined repetition end condition is satisfied;
A parameter estimation apparatus including: The discriminant function is represented by f (w _k ^* ) shown in the following equation:
The parameter estimation apparatus according to claim 1, wherein the objective function is represented by L (A, X, γ) represented by the following expression.
Where w represents the multiple set of data, w ^* represents the multiple set of data to which no category is assigned, and A represents a parameter a _i representing the degree to which each piece of data in the data set contributes to category assignment. Represents a set, 0 ≦ a _i ≦ C, t _i represents a category pre-assigned to the data, X represents a set of latent vectors x _v of each discrete feature v,
Represents a q-dimensional latent vector of the discrete feature v, γ is a parameter for controlling the similarity value between the discrete features, and K (w _i , w _j ; X, γ) is data i and data This is a kernel function that represents the degree of similarity with j. An input unit that accepts an input unknown data set that is a data set of data that can be expressed by multiple sets of discrete features and whose assigned category is unknown,
A latent vector of each discrete feature estimated by the parameter estimation device according to claim 1, a parameter representing a degree of contribution of each of the data of the data set to category assignment, and each of the data of the data set An assigning unit that assigns the category according to the identification function to each data of the unknown data set received by the input unit based on the assigned category and the data set;
Category assignment device including The initialization unit is an input data set to which each category is assigned in advance, and each of the data included in the data set of the data set that can be expressed by a multiple set of discrete features A latent vector of discrete features, a parameter representing the degree to which each of the data of the data set contributes to category assignment, a category assigned to each of the data of the data set, and a category assigned to the data using the data set Initializing a latent vector for each discrete feature and a parameter representing the degree to which each piece of data in the data set contributes to category assignment, used in an identification function for identifying
The latent vector estimator is configured so that each discrete feature is based on a parameter that represents the degree to which each of the data of the data set contributes to category assignment, the category assigned to each of the data of the data set, and the data set. And a parameter representing the degree to which each of the data of the data set contributes to category assignment optimizes the latent vector of each discrete feature such that an objective function representing the degree of effectiveness in categorizing the data is optimized. Estimating, and
The discriminant function estimator optimizes the objective function based on the latent vector of each discrete feature, the category assigned to each of the data of the data set, and the data set. Estimating a parameter representing the degree to which each piece of data contributes to a category assignment;
Repeating the estimation by the latent vector estimation unit and the estimation by the discriminant function estimation unit until the iterative determination unit satisfies a predetermined repetition end condition;
A parameter estimation method including: The discriminant function is represented by f (w _k ^* ) shown in the following equation:
The parameter estimation method according to claim 1, wherein the objective function is represented by L (A, X, γ) represented by the following expression.
Where w represents the multiple set of data, w ^* represents the multiple set of data to which no category is assigned, and A represents a parameter a _i representing the degree to which each piece of data in the data set contributes to category assignment. Represents a set, 0 ≦ a _i ≦ C, t _i represents a category pre-assigned to the data, X represents a set of latent vectors x _v of each discrete feature v,
Represents a latent vector of the q-dimensional discrete feature v, γ is a parameter for controlling the similarity value between the discrete features, and K (w _i , w _j ; X, γ) is data i and data This is a kernel function that represents the degree of similarity with j. An input unit receiving an input unknown data set that is a data set of data that can be expressed by multiple sets of discrete features and whose assigned category is unknown;
An allocation unit includes a latent vector of each discrete feature estimated by the parameter estimation device according to claim 1 and a parameter indicating a degree of each of the data of the data set contributing to category allocation, Assigning the category according to the identification function to each data of the unknown data set received by the input unit based on the category assigned to each of the data and the data set;
Category assignment method including   The program for functioning a computer as each part of the parameter estimation apparatus of Claim 1 or Claim 2, or the category allocation apparatus of Claim 3.