JP5446788B2 - Information processing apparatus and program - Google Patents

Description

  The present invention relates to an information processing apparatus and a program.

  In recent years, when making a proposal such as recommending purchase of a product, a method of making a proposal based on information on other things having the same tendency as the target of the proposal has been used. In this method, for example, when a target person who receives a proposal has purchased a product (product A) in the past, the target person can purchase another product purchased by another person who has purchased the same product A in the past. Make recommendations to recommend. Here, it is considered that the other person who purchased the product A has a relatively high possibility of having the same purchase tendency as the target person compared to the other person who has not purchased the product A. Hereinafter, the process for such a proposal is referred to as a “recommendation process”.

  In particular, in the service based on the long tail theory, the proposal by the above method is considered to be highly effective. The long tail provides a service or the like corresponding to a set composed of a small number that does not form a large-scale set having a large ratio with respect to the whole. Since there are many common points among data (such as purchase histories) included in such a small group, it is highly likely that proposals will be accepted when proposals are made based on differences between data. The For example, there are two customer data B1 and B2 included in a small number of sets, the customer data B1 has purchase history of the product D1 and the product D2, and the customer data B2 has the purchase of the product D1 and the product D3. The product D3 is recommended to the customer corresponding to the customer data B1, and the product D2 is recommended to the customer corresponding to the customer data B2.

  When making a proposal based on data obtained from a group composed of such a small number, conventionally, data extraction for making a proposal by collaborative filtering has been performed (for example, Patent Document 1). Collaborative filtering is a method for extracting another sample data similar to or approximate to a certain sample data based on the cross-correlation from a database including a plurality of sample data, and obtaining the data based on the extracted other sample data It is.

  However, in collaborative filtering, since scanning is based on certain sample data, it has not been possible to know what group is formed by the sample data in the database. For this reason, it has not been possible to perform quantitative data processing on the database, such as analysis based on the tendency of the group made by each sample data included in the database.

  Therefore, a method is known in which clustering is performed by performing analysis processing on a database using an EM algorithm or a variational Bayes method (for example, Patent Documents 2 and 3). Here, clustering is a process of dividing a set of data such as a database into subsets. Each sample data included in the subset obtained by clustering has some common feature, and the feature is different for each subset.

JP 2002-334257 A JP 2007-272291 A JP 2004-117503 A

However, in the clustering disclosed in Patent Documents 2 and 3, the granularity cannot be adjusted. The granularity indicates the scale of the subset.
The clustering granularity will be described with reference to FIG.
The number of sample data S included in one subset shown in the clustering result 101 in FIG. 10 is larger than that in each subset shown in the clustering results 102 and 103. The clustering granularity in such a clustering result 101 is larger than the clustering results 102 and 103.
On the other hand, the number of sample data S included in each subset shown in the clustering result 103 of FIG. 10 is smaller than the subset of the clustering result 101 and the subset 102a of the clustering result 102. The clustering granularity in such a clustering result 103 is smaller than the clustering results 101 and 102.
Thus, the granularity of clustering indicates the number of sample data included in the subset, that is, the scale of the subset.

The granularity of clustering is determined based on a rule relating to common points between sample data included in the subset. For example, when the rules for common points are loose, the granularity tends to increase. A loose rule is, for example, a rule composed of matters with a relatively low probability of occurring in the entire sample data. “A matter with a relatively low probability of occurrence” is a matter where the occurrence frequency is about 0.5 when 0 is assumed when no matter occurs and 1 is assumed when 100%.
Conversely, when the rules for common points are strict, the granularity tends to be small. Strict rules are defined by, for example, the fact that the probability of not occurring in the entire sample data is relatively high (ie, near 0) or the probability of occurrence is relatively high (ie, near 1). Rules.
The above numbers represent the probability of the occurrence. When this probability is close to 1 or 0, the probability that it does not occur is close to 0 or 1, and it is clear whether it will occur or not. On the other hand, if the probability is 0.5, it will be 0.5 if it does not occur. The most ambiguous state.

  In the clustering disclosed in Patent Documents 2 and 3, the granularity of clustering as described above cannot be adjusted.

  An object of the present invention is to make it possible to adjust the granularity of clustering.

The invention according to claim 1 is a storage unit that stores a database including a plurality of sample data, an input unit that inputs a parameter for determining the size of category data, the parameter input by the input unit, and the A control unit that generates one or a plurality of category data, which is one or a plurality of subset data including sample data included in the database, by an algorithm using a mixture distribution model based on a variational Bayes method based on a database; The parameter is a parameter for determining the prior distribution of the mixed distribution and / or the prior distribution on the mixing ratio side in the mixed distribution model, and the mixed distribution model Bernoulli distribution, using Dirichlet distribution as the prior distribution of the mixture ratio side distribution A value obtained by adding 1 to the value of over data on dimensions of the mixture distribution is less than a value obtained by dividing by two, with a beta distribution prior distribution of the mixed Bernoulli distribution side, it sets the value of the parameter to 0.1 or less It is characterized by that.

  A second aspect of the present invention is the information processing apparatus according to the first aspect, wherein the category data has a clustering granularity.

  A third aspect of the present invention is the information processing apparatus according to the second aspect, wherein the parameter is a parameter for determining the size of the clustering granularity.

A fourth aspect of the present invention is the information processing apparatus according to any one of the first to third aspects , wherein the size of the variance of the mixed distribution is the number of category data to be generated and one It has a correlation with the amount of sample data included in the category data.

The invention according to claim 5, based on a computer, means for storing a database including a plurality of sample data, means for inputting a parameter for determining the size of the category data, the input parameter and the database Te, the algorithm using a mixture model variational Bayesian method, means for generating one or more category data which is one or more subsets data including sample data contained in the database, to function as The parameter is a parameter for determining the prior distribution of the mixed distribution and / or the prior distribution on the mixing ratio side in the mixed distribution model, and the mixed distribution model is a mixed Bernoulli distribution The Dirichlet distribution is used for the prior distribution of the mixture ratio side distribution, and A value obtained by adding 1 to the value of the meter to the dimension of the mixture distribution is less than a value obtained by dividing by two, with a beta distribution prior distribution of the mixed Bernoulli distribution side, setting the value of the parameter to 0.1 or less It is characterized by.

  According to the present invention, the granularity of clustering can be adjusted.

It is a block diagram of information processor 1 by the present invention. It is a functional block diagram of information processor 1 when performing clustering processing. It is an example of the contents of the database D. It is a figure which shows an example of a clustering process result. It is a figure which shows another example of a clustering process result. 4 is a flowchart showing a flow of clustering processing of the information processing apparatus 1. It is explanatory drawing which shows extraction of the item by brute force search. It is explanatory drawing which shows an example of extraction of the item R based on the conditional probability P (R | Q) by this embodiment. It is a flowchart which shows the flow of a recommendation process. It is explanatory drawing for demonstrating regarding the granularity of clustering.

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

FIG. 1 shows a block diagram of an information processing apparatus 1 according to the present invention.
The information processing apparatus 1 is a computer and includes a CPU 11, a RAM 12, a ROM 13, a storage device 14, a communication device 15, an input device 16, and a display device 17.

The CPU 11 reads out programs, data, and the like corresponding to the processing contents from the ROM 13 and the storage device 14 and executes them, and performs various processes including operation control of the information processing apparatus 1.
The RAM 12 is a storage device that stores programs, data, and the like read by the CPU 11 and parameters generated by the processing of the CPU 11.

The ROM 13 is a storage device that stores a program, data, and the like read by the CPU 11 in a non-rewritable manner.
The ROM 13 of this embodiment stores a clustering processing program E.

The storage device 14 is a storage device capable of storing a program, data, and the like read by the CPU 11 in a rewritable manner and storing data generated by the processing of the CPU 11 and a program, data, and the like input from the outside. The storage device 14 is configured by, for example, a flash memory, a hard disk drive, another rewritable storage device, or a combination thereof.
The storage device 14 of this embodiment stores a database D including a plurality of sample data.

  The communication device 15 communicates with an external device by connection via a network.

  The input device 16 performs various inputs to the information processing device 1. The input device 16 according to the present embodiment includes input devices such as a keyboard and a mouse, and performs various inputs to the information processing device 1 according to the input operation content of the user. The input device 16 may include a configuration for accepting an input operation by another device or method. For example, character input using a pen, input using a touch panel type input device, or a combination of one or more of these and input using a keyboard, mouse, or the like

  The display device 17 performs various display outputs accompanying the processing of the information processing device 1.

  In the information processing apparatus 1 according to the present embodiment, the CPU 11 reads out the clustering processing program E from the ROM 13 and executes the clustering processing E, thereby realizing functions corresponding to each functional block shown in FIG. 2 to be described later and stored in the storage device 14. Clustering processing based on the contents of the database D is performed.

FIG. 2 shows a functional block diagram of the information processing apparatus 1 when performing clustering processing.
The information processing apparatus 1 functions as an input unit 21, a storage unit 22, a control unit 23, and an output unit 24.

The input unit 21 performs various inputs to the information processing apparatus 1. The function of the input unit 21 depends on the input device 16.
The storage unit 22 stores the database D. The storage unit 22 can store the result of the clustering process. The function of the storage unit 22 depends on a storage device such as the RAM 12, the ROM 13, and the storage device 14.
The control unit 23 reads the database D from the storage unit 22 and performs clustering processing. The function of the control unit 23 is performed by the CPU 11 that executes the clustering processing program E.
The output unit 24 performs various outputs related to the clustering process. The function of the output unit 24 depends on the output device such as the display device 17.

The clustering process by the information processing apparatus 1 according to the present embodiment is based on a mixed distribution model using the variational Bayes method. The mixed distribution model is one of probability models, and the mixed distribution model used in the present embodiment is a mixed Bernoulli distribution.
Hereinafter, the mixed Bernoulli distribution using the variational Bayes method of the present embodiment will be described.

First, the mixed Bernoulli distribution will be described.
The mixed Bernoulli distribution is known as a distribution used for latent class analysis, and is widely used for applications such as binary data clustering and recommendation systems. Furthermore, since the prior distribution of Bernoulli distribution is one hyperparameter as will be described later, it is easy to experiment and analyze. The probability density function of Bernoulli distribution is given by the following equation (1).

Here, x = (x ₁ ,..., X _M ) ^T is data, and μ = (μ ₁ ,..., Μ _M ) ^T and M are the dimensions of parameters and data, respectively. At this time, the mixed Bernoulli distribution is defined by the following equation (2).

（a，b） は事前分布ｐ（π）、ｐ（θ）のパラメータでハイパーパラメータと呼ばれる。 Here, π represents the mixing ratio of B (x | x _k ), and K represents the number of components. Next, a hidden variable for data x is introduced. The hidden variable z is represented as a competitive vector z = (0,..., 1,..., 0) indicating from which distribution the data x originated. In this case, Dirichlet and beta distribution using the Z = a conjugated prior distribution as prior distribution of the distribution of the latent variable z and the data _{x (z 1, ..., z} N), X = (z 1, ..., z N ), Π and θ are given by the following equations (3) to (6), respectively.

(A, b) are parameters of prior distributions p (π) and p (θ) and are called hyperparameters.

  The hyper parameter a is a parameter of the Dirichlet distribution and determines the mixing ratio of the prior distribution. The hyper parameter b is a parameter of the beta distribution, and determines the magnitude of the variance value of the distribution mixed in the prior distribution.

Next, the variational Bayes method will be described.
In this section, Y represents all hidden variables including parameters, and X represents all observable variables. At this time, Equation (7) holds for an arbitrary probability distribution q (Y) and posterior distribution p (Y | X).

ここでＣ１，Ｃ２は正規化定数である。上式は互いの分布による平均値の計算を含んでいる。したがって、変分ベイズ法による学習は（７）および（８）の逐次的繰り返し演算により実行される。

Here, C1 and C2 are normalization constants. The above equation includes the calculation of the mean value from each other's distribution. Therefore, learning by the variational Bayes method is executed by the sequential iterative operations of (7) and (8).

Next, the variational Bayes method of the mixed Bernoulli distribution will be described.
By calculating the general update equations (14) and (15) of the variational Bayes method derived in the previous section under the setting in the description of the mixed Bernoulli distribution described above, the following variational Bayesian learning algorithm VB e of the mixed Bernoulli distribution is given. -step and VB m-step are obtained.
The following formulas (16) and (17) represent VB e-step. The following formulas (18) to (20) represent VB m-step.

以上、式（１６）〜（２２）によるアルゴリズムを実行することで事後分布が得られる。事後分布は、以下の式（２３）、（２４）によって得られる。

Here, it is set as the following formula | equation (21). However, ψ is called a digamma function, and Equation (22) holds.

As described above, the posterior distribution is obtained by executing the algorithm according to the equations (16) to (22). The posterior distribution is obtained by the following equations (23) and (24).

  Note that “learning” refers to arithmetic processing using VB e-step and VB m-step provided by an algorithm such as variational Bayes method. In this example, learning is performed by determining the prior distribution based on the hyperparameters (a, b) that are parameters of the prior distributions p (π) and p (θ).

  Regarding the setting of the hyper parameter, generally, when the hyper parameter a takes a large value, a posterior distribution including more components (for example, sample data) is formed, and the hyper parameter a has a small value. In this case, the number of components for configuring the posterior distribution is narrowed down. As an example of a case where the hyper parameter a takes a small value, the number of categories can be narrowed down by making the hyper parameter a less than or equal to the value obtained by adding 1 to the dimension (M) divided by 2. Has been obtained.

  On the other hand, if the hyper parameter b is reduced, a small category can be found by constructing a group of more probabilistic certainty, and if the hyper parameter b is increased, the smaller category is reversed. There is a tendency that a large tendency can be seen without detection.

  Based on the algorithm shown in the above example, the information processing apparatus 1 performs a clustering process based on the contents of the database D stored in the storage device 14.

FIG. 3 shows an example of the contents of the database D.
The database D shown in FIG. 3 is a database in which the results of answering whether or not each user U1, U2,... “1” is stored in the item that the user answered that he / she wants, and “0” is stored in the item that he / she does not want. Although not shown in FIG. 3, in this example, each user has answered about 83 items, and the total number of users is 1750 users.

FIG. 4 shows an example of the clustering processing result.
The clustering processing result shown in FIG. 4 is an example of the clustering processing result obtained by reading the database D shown in FIG. 3 and using the variational Bayesian learning algorithm based on the mixed Bernoulli distribution using the variational Bayes method. Each category shown in FIG. 4 and FIG. 5 to be described later is a subset composed of a part of the answer results of the users included in the database D. The item on the left side of each category is the item name of the database D shown in FIG. The numerical value on the right side of each category is the user who answered “1” for the item listed on the left side of all the users making up that category, that is, the user who answered “I want that item” The ratio is indicated by 0 to 1 (0 to 100 [%]).

In the clustering processing result shown in FIG. 4, 13 subsets are formed.
In the clustering processing result shown in FIG. 4, among categories 1 to 13, category 1 is the largest subset composed of the most sample data, that is, the number of answer results by the user (number of users). In the following, in categories 2 to 13, subsets with a large number of users constituting each subset follow in ascending order of numbers. That is, category 13 has the smallest number of users.

Each of the categories 11 to 13 is a subset in which there are an extremely large number of users who answered that they wanted “1” for a specific item, that is, that item. For example, the number of users who answered that category 11 wants xbox360 (registered trademark), ps3 (registered trademark), and psp (registered trademark) exceeds 99.99 [%]. Similarly, in category 12, users who want planes and boats (ships) exceed 99.99 [%], and in category 13, users who want 7 items such as “cell phone” exceed 99.99 [%]. ing.
A subset having a very large proportion of users who answered that they want a specific item as shown in categories 11 to 13 is very likely to be a subset by a small number of users having similar preferences and desires. That is, it can be said that the categories 11 to 13 illustrated in FIG. 4 are a subset that reflects the answer results of a small number of users who share a common hobby and desire, that is, a subset that reflects a so-called “minority opinion”.

  When obtaining the clustering processing result shown in FIG. 4, learning is performed on the database D of FIG. 3 with K = 50 for the number of mixture distributions K and a = b = 0.001 for the hyperparameters a and b, and the effective mixture ratio is determined. When categories having an occurrence probability of 0.0001 or more are extracted as categories, a total of 13 subsets including categories 1, 2, 11 to 13 as shown in FIG. 4 can be obtained.

  In this embodiment, b = 0.001, but this value varies depending on the data and the desired granularity. It can be seen from the characteristics of the prior distribution (beta distribution) that it is easy to extract minority opinions by setting the hyper parameter to 1 or less with respect to the prior distribution, but preferably 0.1 or less Is obtained from experiments.

FIG. 5 shows another example of the clustering processing result.
The clustering process result shown in FIG. 5 is an example of the clustering process result obtained under the same conditions as the clustering process result shown in FIG. 4 except for the value of the hyperparameter b. The value set for the hyper parameter b in the learning for obtaining the clustering process result shown in FIG. 5 is larger than the value set for the hyper parameter b in the learning for obtaining the clustering process result shown in FIG. 1.0.

In the clustering processing result shown in FIG. 5, two subsets are formed.
Category 1 shown in FIG. 5 is the ratio of users who answered that they want items corresponding to universal happiness such as house, money, love, friend, job, etc. There are relatively many categories. Category 2 shown in FIG. 5 is the ratio of users who answered based on specific desires such as laptop (laptop computer), ipod (registered trademark), house (shoes), shoes (shoes), and clothes (clothes). There are many categories. In both categories 1 and 2, the percentage of users who answered that they want the item does not reach 40 [%] for all users who make up that category. From this, it can be said that the categories 1 and 2 are large subsets constituted by the answer results of many users as compared to the categories 11 to 13 in FIG.

  In this way, the granularity of clustering can be adjusted by adjusting the values of the hyper parameters a and b.

  In this embodiment, the hyper parameter is directly specified. However, a mechanism for automatically setting the hyper parameter according to various conditions such as the purpose of classification of the user and the number of specified subsets may be provided. Good. The association between the value of the hyper parameter and the condition for setting the hyper parameter is obtained experimentally and is not particularly limited.

Further, the mixed distribution is not limited to the mixed Bernoulli distribution, and other mixed distributions may be used. Other mixed distributions include, for example, a probability model that can be applied to real-valued data such as a mixed normal distribution, a distribution used in latent semantic analysis such as pLSA (Probabilistic Latent Semantic Analysis), and the like.
Further, learning may be performed by a method other than the variational Bayes method. As an algorithm for performing learning, for example, an algorithm that can set a parameter corresponding to variance in advance, such as an EM algorithm, or a technique such as k-means can be given. In the case of the EM algorithm, the hyperparameter is a parameter that determines the variance value of the distribution to be mixed in the mixed distribution model, and the size of the distribution value of the mixed distribution is the number of subset data to be generated and one subset. It has a correlation with the amount of sample data included in the data.

  The clustering processing results obtained in this way can be used for analysis of user preference classifications, as well as various applications such as recommendation systems, etc. that are used as information sources for making proposals to users. Is possible. For example, information that prompts the purchase of an item belonging to the same category based on the clustering processing result of the user's answer result may be delivered to each user constituting the category. Moreover, regardless of whether or not the user is included in the answer result as described above, it is obtained based on the result of the clustering process described above for a user who has performed an action such as purchase or browsing for an item. Information that prompts the user to purchase another item included in the category indicating a high interest in the item in the subset may be delivered to the user. Various other applications are possible.

  If clustering processing is not performed in advance, if there is a user who has performed an action indicating interest, such as purchase, browsing, etc. for a certain item, the same interest for that item after that action, It is necessary to perform the process of extracting the user who performed the action indicating the interest, and then to perform the process such as extracting another item included in the subset highly relevant to the item every time the user action occurs Arise. Such processing requires a high calculation cost, and wastes the processing power of the computer. In addition, the response is poor.

  On the other hand, in the present embodiment, data indicating a subset (subset data) based on the database D can be obtained by performing clustering processing, and the subset data is stored in the storage unit 22 as a subset as necessary. be able to. The subset data handled in the present embodiment is handled as being generated individually for each subset, but generates data for collectively managing information on a plurality of subsets and sample data constituting each subset. You may make it memorize and use.

  The user can obtain information related to the subset based on the database D by reading the subset data, and can analyze and extract data based on the subset. As an analysis of data, for example, an analysis for obtaining information on the relevance of the desire for each item, the correlation of the number of users, and the like. Examples of data extraction include extraction of data related to an item that is closely related to a certain item, extraction of data of a subset that is deeply related to a certain item, and the like.

Hereinafter, the flow of clustering processing of the information processing apparatus 1 will be described with reference to the flowchart shown in FIG.
First, the control unit 23 reads the database D from the storage unit 22 and acquires data to be subjected to clustering processing (step S1).

  Next, parameter determination processing is performed (step S2). The parameter determination process is a process of determining the values of the hyper parameters a and b. For example, the values of the hyper parameters a and b directly input by the user via the input unit 21 or the user via the input unit 21. The values of the hyperparameters a and b determined on the basis of the purpose of the clustering process input in this way, the desired number of subsets, etc.

Next, the control unit 23 performs learning based on the hyper parameters a and b obtained in step S2 (step S3). The algorithm for learning is based on the above description.
Next, the control unit 23 extracts a subset having an effective mixture ratio based on the learning result of step S3, and performs clustering (step S4).
In the present embodiment, the control unit 23 generates subset data indicating the clustering result in step S <b> 4 and outputs and displays it on the output unit 24.

  Next, it is determined whether or not to end clustering (step S5). The determination in step S5 is determined based on whether or not the opinion is extracted based on the clustering result obtained in step S4. For example, when the processing is clustering performed based on a predetermined number of subsets, clustering is completed based on whether or not a subset corresponding to the predetermined number of subsets is obtained, For example, determining whether to perform clustering.

If the clustering is not terminated in step S5 (step S5: NO), the process returns to step S2.
When the clustering is finished in step S5 (step S5: YES), data extraction processing is performed (step S6). The data extraction process is to perform extraction of desired data based on the clustering result obtained in step S4. The data obtained in step S6 includes, for example, data related to items that are closely related to a certain item, data of a subset that is deeply related to a certain item, and the like.

  By the way, for some item (referred to as item Q), an item R with a high conditional probability P (R | Q) is calculated for an item (referred to as item R) that is likely to occur in association with item Q. A specific item R can be obtained.

  Conventionally, when an item R having a high conditional probability P (R | Q) is extracted, a brute force search is performed on the database D after the item Q is obtained as shown in FIG. Items other than the item Q which is data and contained most are extracted as the item R. Such an item R extraction process requires a high calculation cost in order to perform a brute force process, and in addition to wasting the processing power of the computer, the response is also poor.

Therefore, in the present embodiment, the conditional probability P (R | Q) is applied to each subset obtained by the clustering process described above, and the item R having a high conditional probability P (R | Q) is extracted. .
FIG. 8 shows an example of item R extraction based on the conditional probability P (R | Q) according to the present embodiment.
First, the control unit 23 extracts a subset having the highest occurrence probability of the item Q from each subset obtained by the clustering process (step S11). And the control part 23 extracts the item with the highest generation | occurrence | production probability as the item R except the item Q among the items contained in the extracted subset (step S12).

  By applying the conditional probability P (R | Q) to each subset obtained by the clustering process, it is much larger than when brute force processing is performed on all sample data included in the database D. In addition, the calculation cost can be reduced, and the response from obtaining the item Q to obtaining the item R is greatly improved. Therefore, the recommendation such as proposing the item R based on the item Q can be performed at high speed.

  In the case of performing extraction processing based on a plurality of items, such as conditional probability P (R | Q1, Q2) for extracting another item R based on two or more items (for example, items Q1, Q2). The same processing as described above may be performed based on one of the items Q1 and Q2, or an item having a large value of P (R | Q1) · P (R | Q2) may be extracted. Also good.

The flow of the recommendation process for extracting and proposing the item R will be described using the flowchart of FIG.
First, the control unit 23 performs clustering processing (step S21). The clustering process in step S21 is the same as the clustering process described above.
Thereafter, the control unit 23 extracts a subset having the highest occurrence probability of the item Q from the respective subsets obtained by the clustering process (step S22). And the control part 23 extracts the item with the highest generation | occurrence | production probability as the item R except the item Q among the items contained in the extracted subset (step S23). The extraction result of step S23 is displayed on the display unit 24.

And the control part 23 performs the proposal process based on the item R (step S24).
The proposal process based on the item R in step S24 is, for example, a process of providing information for recommending the purchase of the item R to the user who has purchased the item Q. The information provision is performed, for example, by sending an e-mail, sending a direct mail, a catalog, or the like. A control part processes the process relevant to such information provision according to a preset. Preset is a predetermined process. For example, in the case of mail transmission, it is a process of sending a message to a user by adding the item R and information related to the item R to a predetermined template mail. Is a process for outputting the address label of the user, direct mail to be sent to the user, and information indicating the type of catalog to the person in charge of sending.
The contents of the proposal process in step S24, for example, the result of the mail transmission process and other information output are displayed on the display unit 24.

The item R extraction and recommendation process described above is the most generated except for the item Q among the items included in the subset with the highest occurrence probability of the item Q among the subsets obtained by the clustering process. Although it is based on an item with a high probability, the extraction condition is not limited to this.
For example, for items such as items included in each subset, it is possible to extract items in the subset having the lowest parameter (for example, occurrence probability) and the lowest parameters except for the item. In addition, the parameters such as the occurrence probability are not limited to the maximum / minimum, and an item such as an item can be extracted using an arbitrary numerical value or an approximate value thereof.

  As described above, according to the present embodiment, the information processing apparatus 1 generates subset data indicating a subset based on the database D in accordance with the set hyperparameter (a, b). Here, the hyperparameters (a, b) determine the clustering granularity. The user can arbitrarily input and determine the value of the hyperparameter (a, b) via the input unit 21. That is, the user can adjust the clustering granularity by inputting the values of the hyperparameters (a, b). Therefore, the information processing apparatus 1 can adjust the granularity of clustering.

  Furthermore, the hyper parameter a is a parameter of the Dirichlet distribution and determines the mixing ratio of the prior distribution. The hyper parameter b is a parameter of the beta distribution, and determines the magnitude of the variance value of the distribution mixed in the prior distribution. And the control part 23 produces | generates subset data with the algorithm using the variational Bayes method. This makes it possible to adjust the clustering granularity in accordance with changes in the values of the hyperparameters (a, b).

  Further, the magnitude of the variance value of the distribution mixed in the prior distribution determined by the hyperparameter b has a correlation with the number of subset data to be generated and the amount of sample data included in one subset data. As described above, the larger the value of the hyper parameter b, the smaller the number of subsets, and the larger the number of sample data included in one subset, whereas the smaller the value of the hyper parameter b. As the number of subsets increases, the number of sample data included in one subset tends to decrease. Thus, the hyper parameter b functions as a parameter for adjusting the clustering granularity.

  Furthermore, in learning using the mixed Bernoulli distribution, the Dirichlet distribution is used for the prior distribution of the distribution on the mixing ratio side, and the value of the parameter is made equal to or less than the value obtained by adding 1 to the dimension of the mixed distribution divided by 2. By using a beta distribution for the prior distribution on the Bernoulli distribution side and setting the parameter value to 0.1 or less, a small-grained subset suitable for “extraction of minority opinions” can be obtained. As a result, it is possible to analyze and propose data based on a subset that is closely related to a specific item.

Furthermore, the control unit 23 extracts a parameter different from one or more parameters based on one or more parameters included in one or more subsets, such as the above-described recommendation processing. The parameter here is a parameter included in each subset such as an occurrence probability corresponding to an item such as an item name included in the database.
This makes it possible to extract data that is closely related to a certain item without requiring brute force for all sample data in the database, and to quickly extract data required for proposals such as recommendation processing at a low calculation cost. It can be done efficiently.

  Furthermore, by outputting the data extraction result based on the parameters included in the subset and the result of the proposal process to the display unit 24, the result of each process can be clearly indicated to the user. The user can grasp or analyze data from the display content.

  The embodiment of the present invention should be considered that the embodiment disclosed this time is illustrative and not restrictive in all respects. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

For example, various items such as the item name, the number of items, the user's answer, and the number of users in the database D shown in FIG. 3 are merely examples, and can be arbitrarily changed. The number of subsets shown in FIGS. 4, 12, 15 and the like, the items included in each subset, their probabilities, and the flowchart shown in FIG. 6 are merely examples, and the present invention is not limited thereby.
The location of the database and the program read by the control unit is not limited to the above-described embodiment, and may be stored in a medium that can be read by an external storage device or information processing device.

11 CPU
12 RAM
13 ROM
14 storage device 15 communication device 16 input device 17 display device 21 input unit 22 storage unit 23 control unit 24 output unit D database E clustering processing program

Claims (5)

A storage unit for storing a database including a plurality of sample data;
An input unit for inputting parameters for determining the scale of category data;
One or a plurality of subset data including sample data included in the database by an algorithm using a mixture distribution model by a variational Bayes method based on the parameters input by the input unit and the database A control unit for generating category data of
An information processing apparatus comprising:
The parameter is a parameter for determining the prior distribution of the mixed distribution and / or the prior distribution on the mixing ratio side in the mixed distribution model,
The mixed distribution model is a mixed Bernoulli distribution;
The Dirichlet distribution is used as the prior distribution of the distribution on the mixing ratio side, the value of the parameter is made equal to or less than the value obtained by adding 1 to the dimension of the mixed distribution and divided by 2, and the beta distribution is set on the prior distribution on the mixed Bernoulli distribution side. An information processing apparatus characterized in that the parameter value is set to 0.1 or less .   The information processing apparatus according to claim 1, wherein the category data has a clustering granularity.   The information processing apparatus according to claim 2, wherein the parameter is a parameter for determining a size of the clustering granularity. The magnitude of the mixed is the variance of the distribution, any one of claims 1 to 3, characterized in that it has some To correlate the sample data contained in the number of categories the data to be generated and one category data The information processing apparatus described in 1. Computer
Means for storing a database including a plurality of sample data;
It means for inputting parameters for determining a scale of category data,
Based on the input parameter and the database, the algorithm using a mixture model Variational Bayesian method, one or more category data which is one or more subsets data including sample data contained in the database Means for generating,
Is a program for functioning as
The parameter is a parameter for determining the prior distribution of the mixed distribution and / or the prior distribution on the mixing ratio side in the mixed distribution model,
The mixed distribution model is a mixed Bernoulli distribution;
The Dirichlet distribution is used as the prior distribution of the distribution on the mixing ratio side, the value of the parameter is made equal to or less than the value obtained by adding 1 to the dimension of the mixed distribution and divided by 2, and the beta distribution is set on the prior distribution on the mixed Bernoulli distribution side. A program characterized in that the parameter value is set to 0.1 or less .

Images