JP7309673B2 - Information processing device, information processing method, and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

When recommending an item to a user, it is possible for the target user to evaluate the item by referring to the past evaluations of similar users who have similarly evaluated the item evaluated by the target user to whom the item is to be recommended. There are techniques to estimate and complement expensive items. Among such techniques, approaches such as recommending items of one's own business to customers of other businesses are also performed. Here, there is also a technique of assigning a common cluster to a common customer among a plurality of businesses and relating data based on the cluster (for example, see Patent Document 1).

JP 2019-175419 A

The above technique is based on the premise that customers are classified into clusters in advance, and is not a technique for improving the accuracy of clustering. However, since the cluster classification accuracy affects the estimation accuracy, if the cluster classification is fixed, it is considered that there is a limit to the improvement of the estimation accuracy. Thus, it is considered that there is room for improvement in techniques for estimating missing information based on similar information.

The present invention has been made in view of these points, and it is an object of the present invention to provide a technique for improving the estimation accuracy of missing information estimated based on similar information.

A first aspect of the present invention is an information processing apparatus. (1) a user identifier for specifying each user included in a first user group; a plurality of first attribute information included in a first attribute category for each user; first user group information including a plurality of pieces of second attribute information included in the attribute category; (2) user identifiers for specifying each user included in the second user group; and first attribute category for each user (3) a user identifier for identifying each user included in the third user group, and a second attribute category for each user included in the a user group information acquisition unit that acquires third user group information that includes only a plurality of second attribute information, and an affiliation indicating the probability that each user included in the first user group belongs to each of a plurality of clusters a probability setting unit for setting a probability and an attribute probability indicating the realization probability of each of the first attribute information and the second attribute information in each cluster; and based on the belonging probability and the attribute probability, the a complementary data generation unit that generates a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the first attribute information regarding each of the users included in the third user group; identifying whether information generated by the complementary data generating unit is included in the first attribute information and the second attribute information of each user included in the first user group, the second user group, and the third user group; The identification evaluation function determined from the identification function, the generated first attribute information or the second attribute information, and each user included in the first user group, the second user group, and the third user group an updating unit that updates the discriminant function, the belonging probability, and the attribute probability using a generative adversarial network based on a generated evaluation function that evaluates an error of the first attribute information or the second attribute information. , the complementary data generating unit, based on the belonging probability and the attribute probability updated by the updating unit, generates a plurality of the second attribute information about each of the users included in the second user group, and the third user and generating a plurality of pieces of first attribute information about each of the users included in the group.

The probability setting unit includes an initial probability setting unit that sets an initial value of the attribute probability, an affiliation probability updating unit that updates the affiliation probability by the E step of an EM (Expectation-Maximization) algorithm, and an M step of the EM algorithm. an attribute probability updating unit for updating the attribute probabilities by and an optimization unit for causing the attribute probability updating unit to update the attribute probabilities until a predetermined convergence condition is satisfied, and for causing the attribute probability updating unit to update the attribute probabilities; , may be provided.

The first attribute category and the second attribute category may indicate different categories related to products, respectively, and the first attribute information and the second attribute information are the first attribute category and the second attribute, respectively. The information may be information indicating whether or not a product included in the category is purchased, and the information processing device generates the second attribute information included in the second user group generated by the complementary data generation unit. Selecting the product of the second attribute category recommended to each user of the second user group, and the third user based on the first attribute information included in the third user group generated by the complementary data generation unit A recommended product selection unit that selects products of the first attribute category to be recommended to each user of the group may be further provided.

A second aspect of the present invention is an information processing method. In this method, the processor performs (1) a user identifier for specifying each user included in the first user group, a plurality of first attribute information included in the first attribute category for each user, each user (2) a user identifier for identifying each user included in the second user group; Second user group information including only a plurality of first attribute information included in one attribute category; (3) user identifiers for identifying each user included in the third user group; and second attributes related to each user a third user group information containing only a plurality of second attribute information included in the category; and a belonging probability indicating a probability that each user included in the first user group belongs to each of a plurality of clusters. and an attribute probability indicating a realization probability of each of the first attribute information and the second attribute information in each cluster; and based on the belonging probability and the attribute probability, the second user a step of generating a plurality of said second attribute information relating to each user included in the group and a plurality of said first attribute information relating to each user included in said third user group; an identification evaluation function determined from an identification function that identifies whether or not the generated information is included in the first attribute information and the second attribute information of each user included in the second user group and the third user group; The error between the first attribute information or the second attribute information and the first attribute information or the second attribute information of each user included in the first user group, the second user group, and the third user group updating the discriminant function, the membership probabilities, and the attribute probabilities using a generative adversarial network based on a generative evaluation function to be evaluated; and based on the updated membership probabilities and the attribute probabilities, generating a plurality of pieces of second attribute information about each user included in the second user group and a plurality of pieces of first attribute information about each user included in the third user group.

A third aspect of the present invention is a program. This program stores (1) a user identifier for specifying each user included in a first user group, a plurality of first attribute information included in a first attribute category for each user, each user (2) a user identifier for identifying each user included in the second user group; Second user group information including only a plurality of first attribute information included in one attribute category; (3) user identifiers for identifying each user included in the third user group; and second attributes related to each user third user group information containing only a plurality of second attribute information included in the category; and belonging probability indicating the probability that each user included in the first user group belongs to each of a plurality of clusters. and attribute probabilities indicating realization probabilities of the first attribute information and the second attribute information in each cluster, and based on the belonging probabilities and the attribute probabilities, the second user a function of generating a plurality of second attribute information about each user included in the group and a plurality of first attribute information about each user included in the third user group; an identification evaluation function determined from an identification function that identifies whether or not the generated information is included in the first attribute information and the second attribute information of each user included in the second user group and the third user group; The error between the first attribute information or the second attribute information and the first attribute information or the second attribute information of each user included in the first user group, the second user group, and the third user group a function of updating the discriminant function, the membership probabilities, and the attribute probabilities using a generative adversarial network, based on a generative evaluation function to be evaluated; and based on the updated membership probabilities and the attribute probabilities, A function of generating a plurality of pieces of second attribute information about each user included in the second user group and a plurality of pieces of first attribute information about each user included in the third user group is realized.

In order to provide this program or update part of the program, a computer-readable recording medium recording this program may be provided, or this program may be transmitted via a communication line.

Any combination of the above-described components, and expressions of the present invention converted into methods, devices, systems, computer programs, data structures, recording media, etc. are also effective as aspects of the present invention.

According to the present invention, it is possible to improve the estimation accuracy of missing information estimated based on similar information.

FIG. 4 is a diagram for explaining an overview of estimation processing executed by the information processing apparatus according to the embodiment; 1 is a diagram schematically showing a functional configuration of an information processing device according to an embodiment; FIG. It is a figure which shows an example of the initial value of the belonging probability and attribute probability which a probability setting part sets. It is a figure which shows the user database after the complementation by the complementation data production|generation part which concerns on embodiment. It is a figure which shows the user database after completion|finish which performed binarization processing. 4 is a flowchart for explaining the flow of information processing executed by the information processing apparatus according to the embodiment;

<Overview of Embodiment>
An outline of an embodiment will be described. An information processing apparatus according to an embodiment is an apparatus that estimates and complements missing information based on information related to the missing information. As a specific example, the information processing device according to the embodiment performs Based on the purchase histories of a group of users with purchase histories, the tendency of users who have purchase histories of only items included in one category to purchase items included in the other category is estimated.

FIGS. 1(a) and 1(b) are diagrams for explaining an overview of estimation processing executed by the information processing apparatus according to the embodiment. Specifically, FIG. 1A is a diagram for explaining the relationship between user attributes and user groups, and FIG. It is a figure for explaining.

In FIG. 1A, the first user attribute indicates information indicating the purchase history of items included in the first category. Also, the second user attribute indicates the purchase history of items included in a second category different from the first category. In FIG. 1A, the first user group is a set of users having two attributes, a first user attribute and a second user attribute.

A second user group is a set of users who have the first user attribute but do not have the second user attribute. That is, the users included in the first user group are users who cannot grasp the purchasing tendency of items in the second category. On the other hand, the third user group is a group of users who have the second user attribute but do not have the first user attribute. That is, users included in the third user group are users who cannot grasp the purchasing tendency of items in the first category.

FIG. 1(b) is a diagram schematically showing the data structure of a user database that stores a user identifier for specifying each user and attribute information of each user in association with each other. The information processing apparatus according to the embodiment manages each user by uniquely assigning a user identifier. In FIG. 1B, the users included in the first user group are four users assigned user identifiers UID0001, UID0002, UID0003, and UID0004. The users included in the second user group are five users whose user identifiers are UID0005, UID0006, UID0007, UID0008, and UID0009, and the users included in the third user group are two users whose user identifiers are UID0010 and UID0011. is a user of

In FIG. 1(b), the first attribute information includes three pieces of information indicated by 'S1', 'S2' and 'S3', each of which is a purchase order for three different items included in the first category. This is information indicating the presence/absence. Specifically, "1" indicates that there is a purchase, and "0" indicates that there is no purchase. For example, "S1" of a user whose user identification is UID0001 is "1", indicating that this user has purchased an item corresponding to S1 in the past. Similarly, "T1" and "T2" included in the second attribute information are information indicating whether or not two different items included in the second category are purchased. For example, a user whose user identifier is UID0004 has never purchased an item corresponding to T1 in the past.

As indicated by the upward slanted lines in FIG. 1(b), the second attribute information of all the users included in the second user group is blank. This indicates that the second attribute information is unknown for users included in the second user group. In addition, as indicated by the diagonal lines slanting downward to the right in FIG. 1(b), the first attribute information of all the users included in the third user group is blank. This indicates that the first attribute information is unknown for the users included in the third user group. The processing purpose of the information processing apparatus according to the embodiment is to estimate the area indicated by diagonal lines in FIG. That is.

In FIGS. 1(a) and 1(b), for convenience of explanation, the total number of users is 11, but in general the number is often more than 11, and the number of types of attribute information is more than 3. There are many. The information processing apparatus according to the embodiment calculates a statistical relationship between the first attribute information and the second attribute information by analyzing the attribute information of the first user. For example, it is assumed that a user whose first attribute information is all "1" has a high tendency to have "T1" of "1" in the first attribute information. In this case, it is considered highly probable that the users in the second user group whose first attribute information is "1" will purchase the product corresponding to T1 in the second attribute information.

Although the details will be described later, the information processing apparatus according to the embodiment sets two or more clusters, and updates the probability that each user belongs to each cluster using a known EM (Expectation-Maximization) algorithm. presume. The information processing apparatus according to the embodiment further uses a known framework of Generative Adversarial Networks (GAN) to determine whether the estimated information is original information or estimated information The estimation accuracy is improved in the sense that it is difficult to determine whether or not As a result, the information processing apparatus according to the embodiment uses the calculated relationship to improve the estimation accuracy of the second attribute information for the second user group and the first attribute information for the third user group. can.

<Functional Configuration of Information Processing Apparatus 1 According to Embodiment>
FIG. 2 is a diagram schematically showing the functional configuration of the information processing device 1 according to the embodiment. The information processing device 1 includes a storage section 2 and a control section 3 . In FIG. 2, arrows indicate main data flows, and data flows not shown in FIG. 2 may exist. In FIG. 2, each functional block does not show the configuration in units of hardware (apparatus), but the configuration in units of functions. Therefore, the functional blocks shown in FIG. 2 may be implemented within a single device, or may be implemented separately within a plurality of devices. Data exchange between functional blocks may be performed via any means such as a data bus, network, or portable storage medium.

The storage unit 2 includes a ROM (Read Only Memory) for storing the BIOS (Basic Input Output System) of a computer that implements the information processing apparatus 1, a RAM (Random Access Memory) that serves as a work area for the information processing apparatus 1, an OS ( (Operating System), application programs, and a large-capacity storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive) that stores various information referred to when the application program is executed. The storage unit 2 also stores a user database shown in FIG. 1(b).

The control unit 3 is a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) of the information processing device 1, and executes a program stored in the storage unit 2 to acquire the user group information acquisition unit 30, probability It functions as a setting unit 31 , a complementary data generation unit 32 , an update unit 33 and a recommended product selection unit 34 . The probability setting unit 31 includes an initial probability setting unit 310 , a belonging probability updating unit 311 , an attribute probability updating unit 312 and an optimization unit 313 .

Note that FIG. 2 shows an example in which the information processing device 1 is composed of a single device. However, the information processing apparatus 1 may be realized by computational resources such as a plurality of processors and memories, for example, like a cloud computing system. In this case, each unit that configures the control unit 3 is implemented by executing a program by at least one of a plurality of different processors.

The user group information acquisition unit 30 refers to the user database stored in the storage unit 2, and assigns a user identifier for specifying each user included in the first user group and a first attribute category for each user. Acquiring first user group information including a plurality of included first attribute information and a plurality of second attribute information included in a second attribute category for each user. In addition, the user group information acquisition unit 30 refers to the user database, a user identifier for specifying each user included in the second user group, and a plurality of first attributes included in the first attribute category related to each user. Acquire second user group information that includes only attribute information. Further, the user group information acquisition unit 30 refers to the user database, and refers to the user identifier for specifying each user included in the third user group, and the plurality of second attributes included in the second attribute category for each user. Acquire third user group information that includes only attribute information.

The probability setting unit 31 sets the probability that each user included in the first user group belongs to each of a plurality of clusters, and the realization probability of each of the first attribute information and the second attribute information in each cluster. Set the attribute probabilities to indicate

FIGS. 3A to 3C are diagrams showing examples of belonging probabilities and attribute probabilities set by the probability setting unit 31. FIG. Specifically, FIG. 3A is a diagram showing an example of initial values of attribute probabilities set by the probability setting unit 31, and FIG. FIG. 4 is a diagram showing; An initial probability setting unit 310 of the probability setting unit 31 sets initial values of attribute probabilities by using, for example, random numbers.

For example, the probability P that a user whose user identifier is UID00005 belongs to the first cluster (first cluster |ID=UID00005) can be calculated as follows.
P(first cluster|ID=UID00005)∝P(ID=UID00005, first cluster)=P(S1=1|first cluster)×P(S2=1|first cluster)×P(S3=1| first cluster) = 0.89 x 0.0013 x 0.7499 = 0.000896.

Similarly, the probability P that the user whose user identifier is UID00005 belongs to the second cluster (second cluster |ID=UID00005) can be calculated as follows.
P(second cluster|ID=UID00005)∝(ID=UID00005, second cluster)=P(S1=1|second cluster)×P(S2=1|second cluster)×P(S3=1|th 2 clusters) = 0.9856 x 0.0450 x 0.3888 = 0.01723.

The right sides of P (first cluster |ID=UID00005) and P (second cluster |ID=UID00005) are multiplied by the prior probabilities P (first cluster) and P (second cluster) of each class, respectively. is correct, but for the sake of simplicity, it is assumed that P (first cluster)=P (second cluster)=0.5 and is omitted. However, P (first cluster) and P (second cluster) may be estimated without omission.

From this, the posterior probability P (first cluster|ID=UID00005) of the first cluster regarding the user whose user identifier is UID0005 is P (ID=UID00005, first cluster) and P (ID=UID00005, second cluster). can be calculated as follows.

P(first cluster|ID=UID00005)=P(ID=UID00005, first cluster)/(P(ID=UID00005, first cluster)+P(ID=UID00005, second cluster))=0.000896/( 0.000896 + 0.01723) = 0.0494.

On the other hand, the posterior probability P (second cluster |ID=UID00005) of the first cluster regarding the user whose user identifier is UID0005 can be calculated as follows.

P(second cluster|ID=UID00005)=1−P(first cluster|ID=UID00005)=0.9506.

Note that the process of generating the belonging probabilities shown in FIG. 3(b) from the attribute probabilities shown in FIG. 3(a) corresponds to step E in the EM algorithm. The belonging probability updating unit 311 of the probability setting unit 31 updates the belonging probability by executing the E step of the EM algorithm.

The attribute probability updating unit 312 of the probability setting unit 31 updates the attribute probabilities by executing the M step of the EM algorithm based on the belonging probabilities shown in FIG. 3(b). The optimization unit 313 of the probability setting unit 31 causes the belonging probability updating unit 311 to update the belonging probability and the attribute probability updating unit 312 to update the attribute probability until a predetermined convergence condition is satisfied. As a result, the probability setting unit 31 finally generates the attribute probabilities shown in FIG. 3(c). Although not shown, the attribute probability updating unit 312 also calculates final affiliation probabilities based on the attribute probabilities shown in FIG. 3(c).

The example shown in FIG. 3 shows an example in which the probability setting unit 31 sets two clusters, a first cluster and a second cluster. In FIG. 3A, for example, when a user belongs to the first cluster, the possibility that the user has a purchase history of "S3" in the first attribute information is 74.99% (0.7499 ratio). It is shown that. The probability setting unit 31 updates attribute probabilities and belonging probabilities by using the framework of the EM algorithm and GAN.

FIG.3(c) is a figure which shows an example of the attribute probability which the probability setting part 31 produced|generated by update processing. For example, P(T1=1|first cluster) and P(T1=1|second cluster) are calculated as follows.

P(T1=1|first cluster)=
{P(T1=1|ID=UID0001)×P(first cluster|ID=UID0001)
+P(T1=1|ID=UID0002)×P(first cluster|ID=UID0002)
+P(T1=1|ID=UID0003)×P(first cluster|ID=UID0003)
+P(T1=1|ID=UID0004)×P(first cluster|ID=UID0004)
+P(T1=1|ID=UID0010)×P(first cluster|ID=UID0010)
+P(T1=1|ID=UID0011)×P(first cluster|ID=UID0011)}/{P(first cluster|ID=UID0001)+P(first cluster|ID=UID0002)+P(first cluster| ID=UID0003)+P(first cluster|ID=UID0004)+P(first cluster|ID=UID0010)+P(first cluster|ID=UID0011)}
= {1 x 0.5293 + 1 x 0.9970 + 1 x 0.0075 + 0 x 0.0019 + 1 x 0.9558 + 0 x 0.1481} / {0.5293 + 0.9970 + 0.0075 + 0.0019 + 0.9558 + 0.1481}
= 0.9432

P(T1=1|2nd cluster)=
{P(T1=1|ID=UID0001)×P(second cluster|ID=UID0001)
+P(T1=1|ID=UID0002)×P(second cluster|ID=UID0002)
+P(T1=1|ID=UID0003)×P(second cluster|ID=UID0003)
+P(T1=1|ID=UID0004)×P(second cluster|ID=UID0004)
+P(T1=1|ID=UID0010)×P(second cluster|ID=UID0010)
+P(T1=1|ID=UID0011)×P(second cluster|ID=UID0011)}/{P(second cluster|ID=UID0001)+P(second cluster|ID=UID0002)+P(second cluster| ID=UID0003)+P(second cluster|ID=UID0004)+P(second cluster|ID=UID0010)+P(second cluster|ID=UID0011)}
= {1 x 0.4707 + 1 x 0.0030 + 1 x 0.9925 + 0.9981 + 1 x 0.0442 + 0 x 0.8519} / {0.4707 + 0.0030 + 0.9925 + 0.9981 + 0.0442 + 0.8519}
= 0.4495

By the updating process executed by the probability setting unit 31, the initial values of the attribute probabilities shown in FIG. 3(a) are updated to the values shown in FIG. 3(c). When the updating process is repeated, the updated values shown in FIG. 3(c) are set in FIG. 3(a), and the values shown in FIG. 3(c) are updated.

Based on the affiliation probability and the attribute probability, the complementary data generation unit 32 generates a plurality of second attribute information regarding each user included in the second user group and a plurality of first attribute information regarding each user included in the third user group. to generate information; As shown in FIG. 1B, the plurality of second attribute information regarding each user included in the second user group and the plurality of first attribute information regarding each user included in the third user group are both missing. It is information that

For example, the complementary data generation unit 32 can calculate the probability P (T1=1|ID=UID0005) that "T1" is 1 for the user whose user identifier is UID0005 as follows.
P(T1=1|ID=UID0005)=P(T1=1|1st cluster)×P(1st cluster|ID=UID0005)+P(T1=1|2nd cluster)×P(2nd cluster|ID =UID0005) = 0.9432 x 0.0494 + 0.4495 x 0.9506 = 0.4739.

Also, the probability P (T1=0|ID=UID0005) that "T1" is 0 for the user whose user identifier is UID0005 can be calculated as follows.
P(T1=0|ID=UID0005)=1−P(T1=1|ID=UID0005)=1−0.4739=0.5261.

FIG. 4 is a diagram showing the user database after complementation by the complementation data generation unit 32 according to the embodiment. As indicated by the underline in FIG. 4, missing information in the user database shown in FIG.

FIG. 5 is a diagram showing the user database after complementation, and is a diagram showing the user database after complementation that has undergone binarization processing. More specifically, each value in the database shown in FIG. 4 is binarized such that 0 is less than a threshold value (for example, 0.5) and 1 is greater than or equal to the threshold value. In FIG. 5, those for which actual values are given in FIG. 1 fill in the values in FIG.

Returning to the description of FIG. The update unit 33 determines whether information generated by the complementary data generation unit 32 is included in the first attribute information and second attribute information of each user included in the first user group, the second user group, and the third user group. The identification evaluation function determined from the identification function that identifies whether or not, the generated first attribute information or second attribute information, and the first user group, second user group, and third user group of each user included in A generative adversarial network is used to update the discriminant function, membership probabilities, and attribute probabilities based on the generated evaluation function that evaluates the error of the first attribute information or the second attribute information.

The discriminant function is _a function that _{discriminates} whether the data after completion _shown in _FIG . A function that returns 0. A specific example of the identification evaluation function _OD can be expressed by the following equation (1). A larger discriminative evaluation function is better.

An m-dimensional vector input x _. =( _{x.1 , x.2} , . (2) can be expressed.

In equation (2 ₎ , w ₁ , w ₂ , . By passing the function, the output value is in the range [0, 1], and the total output value is 1.

A discriminant function may be learned for each cluster. In that case, _Dc is created for each cluster c, and the identification evaluation function _OD is set as shown in the following equation (3).

Also, as an example, the discriminant function for each cluster c can be expressed as in Equation (4) below.

In Equation (4), w _c1 , w _c2 , . . . , w _cm are the coefficients of the discrimination function for each cluster c. By using a function for each cluster, it is possible to perform discrimination considering the difference in the ratio of the actual value and the complementary value for each cluster. Here, the updating unit 33 repeats learning and updating a predetermined number of times so that the value of _OD increases.

The generation evaluation function _OG is the goodness of complementary generation (reconstruction loss) L(x, n) of the realization value (x _n ∈ X _n ) and the difficulty of authenticity discrimination of the complementary value (x _r ∈ X _r ). It is a function used to evaluate strength. The generation evaluation function _OG is set as a weighted sum of the attribute_S3 and the ease of authenticity discrimination, as shown in the following equation (5).

式（５）において、λは再構成損失と真偽識別の行い易さのバランスをとるハイパーパラメータである。

In Equation (5), λ is a hyperparameter that balances reconstruction loss and ease of authenticity discrimination.

When the discriminant function is learned for each cluster c, the generated evaluation function _OG is set as shown in the following equation (6).

The updating unit 33 repeats learning and updating a predetermined number of times so that the value of the generated evaluation function _OG becomes small. Also, the quality of complementary generation (reconstruction loss) L(x _n ) of the realization value (x _n εX _n ) can be set as shown in Equation (7) below. The smaller the generation evaluation function, the better, and the adversarial generation network optimizes the attribute probabilities and the like through a minimax game in which the generation evaluation function is made smaller while the discrimination evaluation function is increased.

ΣlogL(x _n )=−ΣiεID_S{I(attribute_S331=1, ID_S=i) logP(attribute_S31=1|ID_S=i)+I(attribute_S31=0, ID_S=i) logP(attribute_S31=0 |ID_S=i)+I(attribute_S32=1, ID_S=i) logP(attribute_S32=1|ID_S=i)+I(attribute_S32=0, ID_S=i) logP(attribute_S32=0|ID_S=i)+I (attribute_S33=1, ID_S=i) logP(attribute_S33=1|ID_S=i)+I(attribute_S33=0, ID_S=i) logP(attribute_S33=0|ID_S=i))−ΣjεID_T{I (attribute_T1=1, ID_T=j) logP(attribute_T1=1|ID_T=j)+I(attribute_T1=0, ID_T=j) logP(attribute_T1=0|ID_T=j)+I(attribute_T2=1, ID_T=j) logP(attribute_T2=1|ID_T=j)+I(attribute_T2=0, ID_T=j) logP(attribute_T2=0|ID_T=j)) (7)

In equation (7), ID_S indicates the user identifier of the second user group, and ID_T indicates the user identifier of the third user group. Also, for example, a probability value such as P (attribute_S31=1|ID_S=i) can be obtained by referring to the user database after complement shown in FIG. 4 by the updating unit 33 .

Here, the attribute probability shown in FIG. 3A is a parameter updated in the generated evaluation function, and the complemented user database shown in FIG. It is recalculated from the attribute probabilities shown in FIG. 3(a).

In the above, an example of a simple model that can be calculated only from attribute information is shown, but it may be changed to a complicated model having many parameters related to attribute information, such as a hierarchical Bayesian model. I (attribute S3=1, ID_S=i) is an indicator function that returns 1 if the attribute S3 of ID_S is 1 and returns 0 if it is 0.

Returning to the description of FIG. Based on the belonging probability and the attribute probability updated by the updating unit 33, the complementary data generating unit 32 generates a plurality of pieces of second attribute information about each user included in the second user group and each user included in the third user group. and generating a plurality of first attribute information relating to

In this way, the information processing apparatus 1 according to the embodiment uses the EM algorithm to determine whether each user belongs to each cluster based on the attribute information of the first user group without fixing the cluster to which each user belongs. Estimate probabilities. As a result, compared to the case where the cluster to which each user belongs is fixed, the cluster to which each user belongs can be estimated by statistical analysis of the foresight information.

The information processing apparatus 1 according to the embodiment further uses the values estimated by using the EM algorithm as initial values, so to speak, and uses the GAN framework to improve the accuracy of the estimated loss information. As a result, the information processing device 1 can improve the estimation accuracy of missing information estimated based on similar information.

<Use scene of information processing apparatus 1 according to the embodiment>
The information processing apparatus 1 according to the embodiment selects an item to be recommended to the user, for example, based on the item purchase history of a certain user, from among items included in a category different from the items for which the purchase history exists. can be used for

In this case, it is assumed that the first attribute category and the second attribute category indicate different categories of products. Also, the first attribute information and the second attribute information shall be information indicating whether or not the product included in the first attribute category and the second attribute category is purchased, respectively.

The recommended product selection unit 34 selects products of the second attribute category to be recommended to each user of the second user group based on the second attribute information included in the second user group generated by the complementary data generation unit 32 . Similarly, the recommended product selection unit selects products of the first attribute category to be recommended to each user of the third user group based on the first attribute information included in the third user group generated by the complementary data generation unit 32. . As a result, the information processing apparatus 1 statistically increases the possibility that a user who only has a purchase history of items included in one of the two categories will purchase items included in another category. You can choose expensive items.

<Processing Flow of Information Processing Method Executed by Information Processing Apparatus 1>
FIG. 6 is a flowchart for explaining the flow of information processing executed by the information processing apparatus 1 according to the embodiment. The processing in this flowchart starts, for example, when the information processing apparatus 1 is activated.

The user group information acquisition unit 30 includes (1) a user identifier for specifying each user included in the first user group, a plurality of first attribute information included in the first attribute category related to each user, each user first user group information including a plurality of second attribute information included in the second attribute category relating to each; (2) a user identifier for specifying each user included in the second user group; Second user group information including only a plurality of first attribute information included in the first attribute category; (3) user identifiers for specifying each user included in the third user group; and third user group information including only a plurality of pieces of second attribute information included in the attribute category (step S2).

The probability setting unit 31 sets a belonging probability indicating a probability that each user included in the first user group belongs to each of a plurality of clusters (step S4). Further, the probability setting unit 31 sets attribute probabilities indicating respective realization probabilities of the first attribute information and the second attribute information in each cluster (step S6).

Based on the affiliation probability and the attribute probability, the complementary data generation unit 32 generates a plurality of second attribute information regarding each user included in the second user group and a plurality of first attribute information regarding each user included in the third user group. information is generated (step S8).

The update unit 33 determines whether information generated by the complementary data generation unit 32 is included in the first attribute information and second attribute information of each user included in the first user group, the second user group, and the third user group. The identification evaluation function determined from the identification function that identifies whether or not, the generated first attribute information or second attribute information, and the first user group, second user group, and third user group of each user included in Based on the generated evaluation function that evaluates the error of the first attribute information or the second attribute information, the discriminant function, the belonging probability, and the attribute probability are updated using the adversarial generative network (step S10).

<Effects of Information Processing Apparatus 1 According to Embodiment>
As described above, according to the information processing apparatus 1 according to the embodiment, it is possible to improve the estimation accuracy of missing information estimated based on similar information.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes are possible within the scope of the gist thereof. be. For example, all or part of the device can be functionally or physically distributed and integrated in arbitrary units. In addition, new embodiments resulting from arbitrary combinations of multiple embodiments are also included in the embodiments of the present invention. The effect of the new embodiment caused by the combination has the effect of the original embodiment.

Reference Signs List 1 Information processing device 2 Storage unit 3 Control unit 30 User group information acquisition unit 31 Probability setting unit 310 Initial probability setting unit 311 Belonging probability update Part 312... Attribute probability update part 313... Optimization part 32... Complementary data generation part 33... Update part 34... Recommended product selection part

Claims (5)

(1) A user identifier for specifying each user included in the first user group, a plurality of first attribute information included in the first attribute category for each user, and a second attribute category for each user. (2) a user identifier for identifying each user included in the second user group; and a plurality of items included in the first attribute category for each user (3) a user identifier for identifying each user included in the third user group, and a plurality of second attributes included in the second attribute category for each user a user group information acquisition unit that acquires third user group information that includes only two attribute information;
A belonging probability indicating a probability that each user included in the first user group belongs to each of a plurality of clusters, and an attribute probability indicating a realization probability of each of the first attribute information and the second attribute information in each cluster. and a probability setting unit for setting
Based on the belonging probability and the attribute probability, a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the first attribute information regarding each of the users included in the third user group. and a complementary data generation unit that generates
whether the information generated by the complementary data generation unit is included in the first attribute information and the second attribute information of each user included in the first user group, the second user group, and the third user group; The identification evaluation function determined from the identification function that identifies the generated first attribute information or the second attribute information, and each included in the first user group, the second user group, and the third user group an updating unit that updates the discriminant function, the belonging probability, and the attribute probability using a generative adversarial network based on a generated evaluation function that evaluates an error in the user's first attribute information or second attribute information; with
The complementary data generation unit generates a plurality of pieces of second attribute information about each user included in the second user group and the third user group based on the belonging probability and the attribute probability updated by the updating unit. generating a plurality of said first attribute information for each of the users included in
Information processing equipment. The probability setting unit
an initial probability setting unit that sets an initial value of the attribute probability;
an affiliation probability updating unit that updates the affiliation probability by the E step of an EM (Expectation-Maximization) algorithm;
an attribute probability updating unit that updates the attribute probability according to the M step of the EM algorithm;
an optimization unit that causes the attribute probability updating unit to update the belonging probability until a predetermined convergence condition is satisfied, and causes the attribute probability updating unit to update the attribute probability;
The information processing apparatus according to claim 1, comprising: The first attribute category and the second attribute category indicate different categories of products, respectively,
The first attribute information and the second attribute information are information indicating whether or not the products included in the first attribute category and the second attribute category are purchased, respectively,
The information processing device is
Selecting a product of the second attribute category to be recommended to each user of the second user group based on the second attribute information included in the second user group generated by the complementary data generation unit, and selecting the complementary data further comprising a recommended product selection unit that selects a product of the first attribute category to be recommended to each user of the third user group based on the first attribute information included in the third user group generated by the generation unit;
The information processing apparatus according to claim 1 or 2. the processor
(1) A user identifier for specifying each user included in the first user group, a plurality of first attribute information included in the first attribute category for each user, and a second attribute category for each user. (2) a user identifier for identifying each user included in the second user group; and a plurality of items included in the first attribute category for each user (3) a user identifier for identifying each user included in the third user group, and a plurality of second attributes included in the second attribute category for each user obtaining third user group information including only two attribute information;
A belonging probability indicating a probability that each user included in the first user group belongs to each of a plurality of clusters, and an attribute probability indicating a realization probability of each of the first attribute information and the second attribute information in each cluster. and setting
Based on the belonging probability and the attribute probability, a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the first attribute information regarding each of the users included in the third user group. and generating
From an identification function that identifies whether generated information is included in the first attribute information and second attribute information of each user included in the first user group, the second user group, and the third user group A determined identification evaluation function, the generated first attribute information or the second attribute information, and the first attribute information of each user included in the first user group, the second user group, and the third user group or updating the discriminant function, the membership probabilities, and the attribute probabilities using a generative adversarial network, based on a generated evaluation function that evaluates the error of the second attribute information;
Based on the updated affiliation probability and the attribute probability, a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the second attribute information regarding each of the users included in the third user group. 1 attribute information;
Information processing method that performs to the computer,
(1) A user identifier for specifying each user included in the first user group, a plurality of first attribute information included in the first attribute category for each user, and a second attribute category for each user. (2) a user identifier for identifying each user included in the second user group; and a plurality of items included in the first attribute category for each user (3) a user identifier for identifying each user included in the third user group, and a plurality of second attributes included in the second attribute category for each user a function of acquiring third user group information including only two attribute information;
A belonging probability indicating a probability that each user included in the first user group belongs to each of a plurality of clusters, and an attribute probability indicating a realization probability of each of the first attribute information and the second attribute information in each cluster. and the ability to set
Based on the belonging probability and the attribute probability, a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the first attribute information regarding each of the users included in the third user group. , and a function that generates
From an identification function that identifies whether generated information is included in the first attribute information and second attribute information of each user included in the first user group, the second user group, and the third user group A determined identification evaluation function, the generated first attribute information or the second attribute information, and the first attribute information of each user included in the first user group, the second user group, and the third user group or a function of updating the discriminant function, the belonging probability, and the attribute probability using a generative adversarial network, based on a generated evaluation function that evaluates the error of the second attribute information;
Based on the updated affiliation probability and the attribute probability, a plurality of the second attribute information regarding each of the users included in the second user group and a plurality of the second attribute information regarding each of the users included in the third user group. 1 attribute information;
program to realize

Images