WO2023085279A1 - Information processing system, and information processing method - Google Patents

Abstract

In this information processing system, a first dataset relating to a plurality of first entities is acquired. A second dataset relating to a plurality of second entities is acquired. A group of first feature vectors identified from the first dataset, and a group of second feature vectors identified from the second dataset are subjected to dimensionality reduction processing. As a result, a group of first low-dimensionality feature vectors corresponding to the group of first feature vectors, and a group of second low-dimensionality feature vectors corresponding to the group of second feature vectors are generated. Each first entity is associated with at least one second entity on the basis of the group of first low-dimensionality feature vectors and the group of second low-dimensionality feature vectors.

Description

Information processing system and information processing method Cross-reference to related applications

This international application is Japanese Patent Application No. 2021-182537 filed with the Japan Patent Office on November 9, 2021 and Japanese Patent Application No. 2022 filed with the Japan Patent Office on March 24, 2022. 2022-048893, and the entire contents of Japanese Patent Application No. 2021-182537 and Japanese Patent Application No. 2022-048893 are incorporated by reference into this international application.

The present disclosure relates to an information processing system and an information processing method.

Conventionally, analysis of customer purchasing behavior is performed based on product sales data. Analyzes of customers' contact behavior with respect to mass media and network contents are also performed. A wide variety of information about customers is also collected in the form of questionnaires and face-to-face questions.

Data fusion technology that combines multiple data collected by different means based on common variables is also known. In particular, for a plurality of first customers, a first data set comprising first characteristic data for each customer, and for a plurality of second customers, a second data set comprising second characteristic data for each customer. A technique related to data fusion between sets has already been disclosed (see Patent Document 1, for example). Data fusion combines first characteristic data and second characteristic data of close customers based on variables common between the first data set and the second data set, e.g., demographic attributes of the customers. It is done to bind.

JP 2016-126609 A

With conventional data fusion technology, a common variable regarding customers is required between the first data set and the second data set to be combined in order to determine close customers using the common variable. Therefore, data that do not have common variables cannot be combined.

Therefore, according to one aspect of the present disclosure, the first It is desirable to be able to provide a technique that can realize correspondence between one entity and a second entity.

According to one aspect of the present disclosure, an information processing system is provided. The information processing system includes a first acquisition unit, a second acquisition unit, a dimensionality reduction unit, and an association unit. The first obtaining unit is configured to obtain a first data set for a plurality of first entities. The first data set may describe characteristics of each of the plurality of first entities.

The second acquisition unit is configured to acquire a second data set related to a plurality of second entities. The second data set may describe characteristics of each of the plurality of second entities.

The dimension reduction unit performs dimension reduction processing on a group of first feature vectors identified from the first data set and a group of second feature vectors identified from the second data set, It is configured to generate a first set of low-dimensional feature vectors corresponding to the first set of feature vectors and a second set of low-dimensional feature vectors corresponding to the second set of feature vectors. The group of second low-dimensional feature vectors may be a group of feature vectors having the same number of dimensions as the group of first low-dimensional feature vectors.

Each of the first feature vectors can represent features of a corresponding one of the plurality of first entities. Each of the second feature vectors may represent features of a corresponding one of the plurality of second entities.

The associating unit associates each of the plurality of first entities with at least one of the plurality of second entities based on the group of first low-dimensional feature vectors and the group of second low-dimensional feature vectors. configured to match.

between the first feature vector and the second feature vector if the first set of entities and the second set of entities are subsets from a common population or a mutually related population Even in the absence of common variables, dimensionality reduction allows features of the first entity and features of the second entity to be represented by combinations of components that are common or related to each other.

That is, according to dimensionality reduction, it is possible to extract main feature components that are common or related to each other from the first feature vector and the second feature vector. Therefore, by comparing the low-dimensional feature vectors, it is possible to appropriately determine the degree of matching between the first entity and the second entity.

Therefore, according to one aspect of the present disclosure, the first It is possible to properly associate one entity with a second entity.

According to one aspect of the present disclosure, the associating unit is the similarity between the first entities identified from the first group of low-dimensional feature vectors and the second group of low-dimensional feature vectors identified from the Each of the plurality of first entities is divided into a plurality of It can be associated with at least one of the second entities.

If the first set of entities and the second set of entities are subsets from a common population or mutually related populations, then the mutual relationships in terms of similarity between the entities are similar to the populations. In addition, the first set of entities and the second set of entities generally have or are related to each other.

Therefore, each of the plurality of first entities is associated with at least one of the plurality of second entities such that the correlation between the first entities in terms of similarity matches the correlation between the second entities. Possibly, each first entity can be associated with a suitable second entity that is highly identical or closely related.

According to one aspect of the present disclosure, a group of first low-dimensional feature vectors may be defined by a first feature space. A second set of low-dimensional feature vectors may be defined by the second feature space.

The associating unit is configured such that the distribution of the plurality of first entities in the first feature space specified from the group of first low-dimensional feature vectors is specified from the group of second low-dimensional feature vectors. A mapping may be searched to map the plurality of first entities on the first feature space to the second feature space to match the distribution of the plurality of second entities in the feature space.

The associating unit may be configured to associate each of the plurality of first entities with at least one of the plurality of second entities based on the mapping.

　According to one aspect of the present disclosure, the associating unit is an expression including matrix K, matrix L, and matrix H

A matrix Ω that maximizes the value Z(Ω) according to is searched as a matrix Ω ^* , and based on the matrix Ω ^* , each of the plurality of first entities is associated with at least one of the plurality of second entities. can be configured to T is the transpose symbol. trace is the diagonal sum of matrix X;

Matrix K can be a matrix with N rows and N columns. The number of first entities may be N; The number of second entities can be the same as the first entity. The matrix K is a first similarity matrix in which the value of the element in the i-th row and j-th column represents the similarity between the i-th entity and the j-th entity among the plurality of first entities. obtain.

The value of the i-th row and j-th column element in the matrix K is the first low-dimensional feature vector of the i-th entity among the plurality of first entities and the j-th and a first low-dimensional feature vector of the entity.

The matrix L can be a matrix with N rows and N columns. The matrix L is a second similarity matrix in which the value of the element in the i-th row and j-th column represents the similarity between the i-th entity and the j-th entity among the plurality of second entities. .

The value of the i-th row and j-th column element in the matrix L is the second low-dimensional feature vector of the i-th entity among the plurality of second entities and the j-th feature vector of the plurality of second entities. and a second low-dimensional feature vector of the entity.

The matrix H can be a matrix with N rows and N columns. The matrix H may be a matrix in which the value of the element in the i-th row and j-th column indicates the value 1−1/N when i=j, and indicates the value 0 when i≠j.

According to one aspect of the present disclosure, the associating unit may change the dimensionality reduction method in the dimensionality reduction process based on the matrix Ω ^* . For example, the associating unit, of the group of first low-dimensional feature vectors and the group of second low-dimensional feature vectors, between the first low-dimensional feature vector and the second low-dimensional feature vector corresponding to each other The dimension reduction method in the dimension reduction process may be changed so that the distance in the feature space of is shortened.

According to one aspect of the present disclosure, the associating unit improves the matrix Ω* by repeatedly performing a re-search process on the matrix Ω ^* until a predetermined condition is satisfied, and converts the improved matrix ^{Ω* to} Based on this, each of the plurality of first entities may be configured to correspond to at least one of the plurality of second entities.

The re-searching process may include changing the dimensionality reduction scheme in the dimensionality reduction process based on the matrix Ω ^* . In the re-search process, the dimensionality reduction unit executes the dimensionality reduction process in the dimensionality reduction method after the change, and the group of the first low-dimensional feature vectors and the second low-dimensional feature vector newly obtained thereby. It may involve re-searching the matrix Ω ^* based on the constellation.

According to the information processing system in which the association unit is configured in this way, the association between the first entity and the second entity can be performed with high accuracy.

According to one aspect of the present disclosure, the first data set may include multiple pieces of first feature data. Each of the plurality of first feature data can represent features of a corresponding one of the plurality of first entities. The second data set may include multiple second feature data. Each of the plurality of second feature data can represent features of a corresponding one of the plurality of second entities.

According to one aspect of the present disclosure, the information processing system may further include a data fusion unit. The data fusion unit adds the plurality of second feature data to each of the plurality of first feature data based on the association between the plurality of first entities and the plurality of second entities by the association unit. Combining one of them may be configured to generate an augmented data set. An extended data set may comprise multiple extended data. Each of the plurality of extended data can be combined data of corresponding one first feature data and second feature data.

According to such an information processing system, it is possible to generate a data set with a large amount of information by combining multiple data sets.

According to one aspect of the present disclosure, the first entity may be a person. A second entity can be a person. The first data set may be a data set describing a first characteristic of each of a plurality of persons belonging to the first population. The second data set can be a data set describing a second characteristic of each of the plurality of persons belonging to the second population.

It is believed that features related to people's behavior and interests greatly affect demographic attributes, and that feature distributions according to demographic attributes do not change significantly even among different groups of people. Therefore, according to the information processing system according to one aspect of the present disclosure, it is possible to appropriately associate people in different groups without common variables.

According to one aspect of the present disclosure, the combination of the first feature and the second feature includes a feature related to purchasing behavior, a feature related to movement in at least one of the online space and the offline space, and/or spatially and features relating to visits to multiple points of. Entity matching based on datasets related to these features, and furthermore data fusion, aids in human behavior analysis.

According to one aspect of the present disclosure, the second data set may be associated with identification information of information terminals corresponding to each of the plurality of second entities.

According to one aspect of the present disclosure, an information processing system includes at least part of a set of second entities, among a plurality of second entities, that are associated with any one of a plurality of first entities by an associating unit. as a distribution destination of the information content.

According to one aspect of the present disclosure, the information processing system includes a distribution unit configured to distribute information content to a set of information terminals corresponding to distribution destinations of the information content based on the identification information. may

This information processing system functions meaningfully when the first entity and the second entity are people. According to the distribution method described above, even when the relationship between the first entity and the information terminal is unknown, the identification information of the information terminal associated with the second entity is used to support the first entity. The information content can be appropriately distributed to the information terminal of the second entity that

According to one aspect of the present disclosure, the selection unit includes a first set that is a set of second entities associated with any of the plurality of first entities by the association unit, and a plurality of second entities. Of these, the second set having similar features to the first set may be selected as the distribution destination of the information content. According to such selection of distribution destinations, it is possible to distribute the information content by expanding the distribution destinations within an appropriate range based on the second data set.

According to one aspect of the present disclosure, the second data set may be a data set describing behavioral characteristics of each of the plurality of second entities. In this case, the information processing system may include an estimation unit that calculates, for each of the one or more attention entities, an estimated value regarding the behavior of the corresponding attention entity. One or more entities of interest may be at least a portion of the plurality of first entities. The estimated value may be calculated based on at least one behavioral feature of the plurality of second entities associated with the corresponding entity of interest. The first entity and the second entity can be people.

According to the information processing system including the estimation unit described above, it is possible to estimate the behavior of the first entity through the second data set, which cannot be determined by the first data set alone. An estimate may be a prediction.

According to one aspect of the present disclosure, an information processing method corresponding to the method executed by the information processing system described above may be provided. According to one aspect of the present disclosure, a computer-implemented information processing method may be provided. The information processing method may include obtaining a first data set relating to a plurality of first entities, the first data set describing characteristics of each of the plurality of first entities.

The information processing method may include obtaining a second data set relating to a plurality of second entities, the second data set describing characteristics of each of the plurality of second entities.

The information processing method performs dimension reduction processing on a group of first feature vectors identified from the first data set and a group of second feature vectors identified from the second data set, It may include generating a first set of low-dimensional feature vectors corresponding to the first set of feature vectors and a second set of low-dimensional feature vectors corresponding to the second set of feature vectors. The group of second low-dimensional feature vectors may be a group of feature vectors having the same number of dimensions as the group of first low-dimensional feature vectors.

Each of the first feature vectors can represent features of a corresponding one of the plurality of first entities. Each of the second feature vectors may represent features of a corresponding one of the plurality of second entities.

The information processing method converts each of the plurality of first entities into at least one of the plurality of second entities based on the set of first low-dimensional feature vectors and the set of second low-dimensional feature vectors. may include matching.

According to one aspect of the present disclosure, the matching is a measure of similarity between a first entity identified from a first set of low-dimensional feature vectors and a measure of similarity between entities identified from a second set of low-dimensional feature vectors. Based on the similarity between the second entities, each of the plurality of first entities is combined with the plurality of second entities such that the correlation between the first entities with respect to the similarity matches the correlation between the second entities. It may involve mapping to one of two entities.

According to this information processing method, as in the information processing system described above, a first data set regarding a plurality of first entities and a second data set regarding a plurality of second entities are obtained without using a common variable. A correspondence between the first entity and the second entity can be realized based on .

According to one aspect of the present disclosure, a computer program including instructions for causing a computer to execute the information processing method described above may be provided. According to one aspect of the present disclosure, a computer-readable non-transitory tangible recording medium storing a computer program may be provided.

1 is a block diagram showing the configuration of an information processing system; FIG. 4 is a flowchart representing analysis processing executed by a processor; FIG. 3A is a diagram illustrating the configuration of the first data set, and FIG. 3B is a diagram illustrating the configuration of the second data set. 4A and 4B are diagrams illustrating a search method for the matrix Ω. 4 is a diagram illustrating the configuration of a correspondence table generated by a processor; FIG. FIG. 4 is a diagram illustrating the configuration of an extended data set generated by a processor; It is a flow chart showing analysis processing which a processor performs in a second embodiment. It is a flow chart showing evaluation processing which a processor performs in a third embodiment. FIG. 11 is a flowchart showing selection processing executed by a processor in the third embodiment; FIG. It is a block diagram showing the structure of the delivery system of 4th embodiment. It is a figure which illustrates the structure of the internal data set in 4th embodiment. It is a flow chart showing distribution control processing which a processor performs in a fourth embodiment. It is a flow chart showing distribution control processing which a processor performs in a fifth embodiment. FIG. 16 is a flowchart showing prediction processing executed by a processor in the sixth embodiment; FIG.

REFERENCE SIGNS LIST 1 information processing system 11, 31 processor 13, 33 memory 15, 35 storage 15A first data set 15B second data set 15C extended data set 17 user interface , 19, 39... communication interface, 30... delivery system, 35A... external data set, 35B... internal data set, 35C... extended data set, 40... user company side system, Pr, Pr1... computer program.

Exemplary embodiments of the present disclosure will be described below with reference to the drawings.

<First embodiment>
The information processing system 1 of this embodiment is configured by installing a dedicated computer program Pr in a general-purpose computer. The information processing system 1 includes a processor 11, a memory 13, a storage 15, a user interface 17, and a communication interface 19, as shown in FIG.

The processor 11 executes processing according to the computer program Pr stored in the storage 15. The memory 13 is a primary storage device having a RAM, and is used as a work area when the processor 11 executes processing.

The storage 15 is a secondary storage device including, for example, a hard disk drive or a solid state drive, and stores various data provided during execution of processing according to the computer program Pr in addition to the computer program Pr.

The user interface 17 includes an input device and a display. The input device is provided for inputting an operation signal from a user who operates the information processing system 1 to the processor 11 . A display is provided for displaying various information to the user. Examples of input devices include keyboards and pointing devices.

The communication interface 19 includes a LAN (Local Area Network) interface and a USB (Universal Serial Serial) interface, and is used for communication with external devices. The information processing system 1 transmits and receives data to and from an external device through the communication interface 19 .

The processor 11 in the information processing system 1 executes a process according to the computer program Pr to extend the first data set 15A acquired from the external device through the communication interface 19 using the second data set 15B to obtain an extended data set 15C. to generate

The extended data set 15C is a data set obtained by adding information provided in the second data set 15B to the first data set 15A. Expansion increases the amount of information for each entity that the first data set 15A describes. An entity is, for example, a person, in particular an individual. The increase in the amount of information is performed for human behavior analysis and advertisement distribution based on the extended data set 15C.

Specifically, when an execution command is input from the user through the user interface 17, the processor 11 of the information processing system 1 executes the analysis process shown in FIG. When the analysis process shown in FIG. 2 is started, the processor 11 acquires the first data set 15A and the second data set 15B for data fusion (S110, S120).

In S110 and S120, the processor 11 can read the first data set 15A and the second data set 15B pre-stored in the storage 15 from the storage 15. Thereby, the processor 11 can acquire the first data set 15A and the second data set 15B.

The first data set 15A and the second data set 15B to be acquired can be specified by the user. The user can collect the first data set 15A and the second data set 15B for data fusion in advance and store them in the storage 15 .

Alternatively, the processor 11 can acquire the first data set 15A from the first external device and the second data set 15B from the second external device through communication using the communication interface 19.

A first data set 15A is a data set relating to a plurality of first entities and a data set describing the first characteristics of each of the first entities. The first data set 15A is a set of first feature data, each piece of first feature data representing a first feature of a corresponding one of the plurality of first entities.

A second data set 15B is a data set relating to a plurality of second entities and a data set describing the second characteristics of each of the second entities. The second feature can be a different feature than the first feature. Specifically, the second data set 15B is a set of second feature data, and each of the second feature data is the second feature of a corresponding one of the plurality of second entities. represents

The first set of entities and the second set of entities are, for example, different subsets of a common population. A population can be a collection of people or a collection of consumers. For example, a first set of entities may be a set of people corresponding to customers of the first business. For example, the second set of entities may be a set of people corresponding to customers of a second business that is different from the first business.

Alternatively, the first set of entities may be a set of people whose first actions are collected. The second set of entities may be a set of people from whom the second behavior is collected.

A first data set 15A shown in FIG. 3A is data relating to a first group of people, and includes feature data relating to purchasing behavior of each person. Each piece of feature data is associated with a corresponding person's ID, and indicates by a binary value of 1 or 0 whether the corresponding person has purchased each of the plurality of products P1, P2, P3, .

A second data set 15B shown in FIG. 3B is data relating to a second set of people and comprises feature data relating to browsing behavior of web content for each person. Each feature data is associated with a corresponding person's ID, and whether or not the corresponding person has visited each of the plurality of websites S1, S2, S3, . . . is represented by a binary value of 1 or 0.

In S110, the processor 11 generates an M1-dimensional feature vector x=(x1, x2, x3,...). According to one example, the elements x1, x2, x3, .

Similarly, in S120, the processor 11 calculates the M2-dimensional feature vector y=(y1 , y2, y3, . . . ). According to one example, the elements y1, y2, y3, .

After that, the processor 11 performs dimension reduction processing (S130) on a group of feature vectors x to convert each feature vector x from the M1-dimensional feature vector to a lower-dimensional feature vector Dx=(Dx1 , Dx2, . . . ). Thereby, the processor 11 generates a group of low-dimensional feature vectors Dx corresponding to the group of feature vectors x. The lower right area of FIG. 3A shows an example of a low-dimensional feature vector Dx in the form of a table.

Further, the processor 11 performs dimension reduction processing (S140) on a group of feature vectors y to convert each feature vector y from the M2-dimensional feature vector to a low-dimensional feature vector Dy=(Dy1, Dy1, which is a smaller M-dimensional feature vector). Dy2,...). Thereby, the processor 11 generates a group of low-dimensional feature vectors Dy corresponding to the group of feature vectors y. The low-dimensional feature vector Dy is a feature vector having the same dimension number M as the low-dimensional feature vector Dx. The lower right area of FIG. 3B shows an example of a low-dimensional feature vector Dy in the form of a table.

Examples of algorithms for realizing mapping to a low-dimensional space include nonnegative matrix factorization, latent dirichlet allocation, singular value decomposition, and stochastic Latent semantic analysis (Probabilistic Latent Semantic Analysis) is known. The dimensionality reduction process at S130, S140 can be performed using one of these algorithms.

According to the algorithm described above, the feature vector can be reduced in dimension so that the main feature components that strongly characterize the individual entity are extracted. Alternatively, the feature vector can be reduced in dimension in a form that is less lossy of information to distinguish individual entities.

After that, the processor 11 performs alignment processing to calculate the relationship between the first entity and the second entity based on the set of low-dimensional feature vectors Dx and the set of low-dimensional feature vectors Dy ( S150-S180).

Alignment processing is performed using kernelized sorting technology. Details of alignment processing using kernelized sorting will be described below. However, the alignment process may be realized using adversarial learning, Gromov-Wasserstein Alignment technology, or Unbalanced Optimal Transport technology.

At S150, the processor 11 uses the group of low-dimensional feature vectors Dx to generate a similarity matrix K for the first set of entities. The similarity matrix K is a square matrix with N rows and N columns. Here, N is the number of low-dimensional feature vectors Dx, in other words, the number of first entities.

The similarity matrix K is defined as a matrix in which the value Kij of the i-th row and j-th column element represents the similarity between the i-th entity and the j-th entity in the first entity set.

That is, the similarity matrix K is defined as a matrix that describes the distribution of similarities between entities with respect to the first set of entities. In other words, the similarity matrix K is defined as a matrix that describes the distribution of entities on the feature space with respect to the first set of entities using a measure of closeness between entities.

Specifically, the similarity is calculated using a low-dimensional feature vector Dx[i], which is the low-dimensional feature vector Dx of the i-th entity, and a low-dimensional feature vector Dx[j], which is the low-dimensional feature vector Dx of the j-th entity. and are substituted into the kernel function k(a, b) as a value k(Dx[i], Dx[j]). That is, Kij=k(Dx[i], Dx[j]).

An example of the kernel function k(a,b) includes a Gaussian RBF (radial basis function) kernel expressed by the following equation. The similarity calculated using this kernel function k(a, b) takes values ranging from 0 to 1.

According to the above kernel function k(a, b), the value Kij of the elements of the similarity matrix K is 0<Kij≤1.

At S160, the processor 11 uses the group of low-dimensional feature vectors Dy to generate a similarity matrix L for the second set of entities. The similarity matrix L is a square matrix with N rows and N columns. Here, N is the number of low-dimensional feature vectors Dy, in other words, the number of second entities. That is, the number of first entities and the number of second entities are the same.

In the similarity matrix L, similar to the similarity matrix K, the value Lij of the element in the i-th row and j-th column indicates the similarity between the i-th entity and the j-th entity in the set of second entities. is defined as a matrix representing That is, the value Lij of the element in the i-th row and j-th column is Lij=k(Dy[i], Dy[j]).

In subsequent S170, processor 11 uses similarity matrix K and similarity matrix L to search for matrix Ω that maximizes value Z(Ω) according to the following equation as matrix Ω ^* .

Here, the matrix H is a matrix of N rows and N columns, and the value of the element in the i-th row and j-th column indicates the value 1-1/N when i = j, and when i ≠ j It is a diagonal matrix showing the value 0. T is the transpose symbol. trace(X) is the diagonal sum of matrix X; Similarity matrices K and L are symmetric matrices. The value Z(Ω) is maximized when the ideal Ω is found such that the matrix Ω ^T L'Ω is the transpose of the matrix K'.

Searching the matrix Ω ^* is the similarity between a first entity identified from a set of low-dimensional feature vectors Dx and the similarity between a second entity identified from a set of low-dimensional feature vectors Dy each of the plurality of first entities to at least one of the plurality of second entities such that the correlation between the first entities in terms of similarity matches the correlation between the second entities based on Corresponding to correspond.

In other words, searching the matrix Ω ^* is the distribution of the first entity in the first M-dimensional feature space identified from the group of low-dimensional feature vectors Dx, defined by the similarity between the entities. A plurality of to a second M-dimensional feature space.

The left graph in FIG. 4A conceptually represents the distribution of the first entity, and the left graph in FIG. 4B conceptually represents the distribution of the second entity. The examples shown in FIGS. 4A and 4B define two-dimensional low-dimensional feature vectors Dx and Dy for technical explanation only. Each point labeled E11, E12, E13, E14, E15, E16, E17 indicates the position of the first entity on the feature space. Each point labeled E21, E22, E23, E24, E25, E26, E27 indicates the position of the second entity on the feature space.

As can be understood from FIG. 4B, according to this example, the component Dy1 of the low-dimensional feature vector Dy corresponds to the component Dx2 of the low-dimensional feature vector Dx, and the component Dy2 of the low-dimensional feature vector Dy corresponds to the low-dimensional feature vector It corresponds to the component Dx1 of Dx.

That is, according to the example shown in FIG. 4A , the first group of entities and the second group of entities are such that the entity arrangement and dimension order are between the similarity matrix K and the similarity matrix L. They represent similarity distributions for sets of entities that are essentially the same, only defined differently.

Low dimensionality of feature vectors x, y when the first group of entities and the second group of entities have collective properties that are common or related to each other, such as because the populations are the same By the transformation, even if there is no common variable between the first data set 15A and the second data set 15B of the information sources, it is possible to extract the essential common feature components for each entity.

However, even with such dimensionality reduction, the low-dimensional feature vectors Dx and Dy only have the same feature component, and the arrangement of the feature components cannot be aligned. Also, the entities are not aligned between the first data set 15A and the second data set 15B.

The search of the matrix Ω ^* corresponds to the work of searching for the correspondence between irregular feature vectors Dx and Dy with respect to the array of entities and the array of dimensions, using the identity of the similarity distribution as a clue.

At subsequent S180, the processor 11 associates each of the first entities with at least one of the second entities based on the matrix Ω ^* . According to the similarity distribution, the element value of the i-th row and j-th column of the matrix Ω ^* is the i-th entity in the first entity set and the j-th entity in the second entity set. and represent the degree or possibility of correspondence.

Each element of the matrix Ω ^* ideally takes 0 or 1, the sum of the element values in one row is 1 for each row, and the sum of the element values in one column is 1 for each column. When the matrix Ω ^* is such an ideal matrix, the first entity of the row number and the second entity of the column number of the 1-valued elements correspond to each other.

That is, when the i-th row and j-th column element in the matrix Ω ^* has the value 1, the i-th entity in the first set of entities and the j-th entity in the second set of entities , correspond to each other.

However, in numerical calculations, the matrix Ω ^* is rarely such an ideal matrix. Therefore, in S180, each of the plurality of first entities is associated with at least one of the second entities using one of the following methods.

(Method 1) The i-th row of the matrix Ω ^* is searched for the element with the maximum value. If the element with the largest value is in the c-th column, the i-th entity in the first set of entities is associated with the c-th entity in the second set of entities. Do this for all rows.

With this method, one of the second entities may be associated with multiple first entities. To limit this possibility, a neighborhood search may be performed. Contextual dissimilarity measure is known as an example of neighborhood search.

(Method 2) In order to perform a strict one-to-one correspondence, by solving an optimal assignment problem with the matrix Ω ^* as an input, each of the plurality of first entities is assigned to one of the non-overlapping second entities. correspond to

At S180, the processor 11 can further output the correspondence table shown in FIG. 5 as a table describing the correspondence between the first entity and the second entity. That is, a correspondence table describing the ID of the corresponding second entity in association with each ID of the first entity can be output and stored in the storage 15 .

Further, the processor 11 executes data fusion processing (S190). In the data fusion process, the processor 11 combines the first data set 15A and the second data set 15B based on the correspondence result or the correspondence table to generate the extended data set 15C. .

The extended data set 15C comprises multiple extended data. As shown in FIG. 6, each of the plurality of extension data is combined data of corresponding one first feature data and second feature data.

That is, the processor 11 assigns each of the plurality of first feature data included in the first data set 15A to each of the plurality of second feature data included in the second data set 15B based on the correspondence table. Combining one produces an extended data set 15C.

When the correspondence table associates the i-th entity in the first entity set with the j-th entity in the second entity set, the processor 11 of the i Generate extended data for the th entity.

The extended data set 15C generated in this way is stored in the storage 15. The extended data set 15C stored in the storage 15 is transferred to another system through the communication interface 19 based on a command from the user input through the user interface 17, for example.

Another system may be, for example, an advertisement distribution system. Based on the extended data set 15C, the advertisement distribution system can determine the entity to which the advertisement is to be distributed, and distribute the advertisement to the entity.

At S190, when the data fusion process ends, the processor 11 ends the analysis process shown in FIG.

As described above, according to the information processing system 1 of the present embodiment, even if there is no common variable between the first data set 15A and the second data set 15B, the similarity distribution Based on this, the first entity and the second entity can be appropriately associated.

For proper matching, it is preferable that the similarity distributions between the first set of entities and the second set of entities match, are similar to, or are related to each other.

Such favorable conditions are approximately satisfied when the first set of entities and the second set of entities are subsets from the same population. Therefore, when the first entity and the second entity are people, that is, when data sets representing features related to people are handled as the first data set 15A and the second data set 15B, Technology works meaningfully.

In particular, human behavior often shows trends according to demographic attributes. Therefore, when the first data set 15A and the second data set 15B are data sets based on collected data from populations whose distributions of demographic attributes are estimated to be similar to each other, appropriate matching between entities is feasible.

For example, the first data set 15A and the second data set 15B are data sets that explain the characteristics of people belonging to different groups with no common variables, or data sets that explain the characteristics of different behaviors. Even if there is, the correspondence between the entities can be made appropriately. Therefore, it is possible to generate a data set useful for human psychology/behavior analysis as the extended data set 15C.

According to the example described above, the first data set 15A is a data set describing the characteristics of purchasing behavior of each of a plurality of people belonging to the first group, and the second data set 15B is a A data set describing characteristics of website visit behavior and/or web content browsing behavior of each of a plurality of people belonging to a group.

According to another example, one of the first data set 15A and the second data set 15B may be a data set that describes characteristics of a person's media contact behavior, such as television viewing behavior. One of the first data set 15A and the second data set 15B may be a data set that describes the characteristics of usage of a mobile terminal such as a smart phone.

A dataset that describes the characteristics of a person's movement in an offline space (that is, a real space) may be used as one of the first dataset 15A and the second dataset 15B. A dataset may describe features of a person's movement in the offline space, such as visits to multiple locations, travel routes, and/or means of travel.

A dataset that describes the characteristics of people's movement in the online space may be used as one of the first dataset 15A and the second dataset 15B. A dataset may describe features of a person's movement and surfing in a virtual reality (VR) space as features of a person's movement in an online space. A data set based on data collected by a questionnaire may be used as one of the first data set 15A and the second data set 15B.

As a combination of the first data set 15A and the second data set 15B, a combination of a data set collected by questionnaire and a data set related to TV viewing behavior, or a data set related to movement history and a data set related to purchase may be employed in combination with

In the above embodiment, processing such as ZCA whitening, normalization, and standardization may be performed on the group of low-dimensional feature vectors Dx and Dy.

In the above embodiment, the number of dimensions M of the low-dimensional feature vectors Dx and Dy is determined by the designer or user, but the information processing system 1 may be configured to search for the optimum number of dimensions M. For example, the information processing system 1 repeatedly executes the analysis processing shown in FIG. It may be configured to automatically select the number M.

<Second embodiment>
The information processing system 1 of the second embodiment is configured such that the processor 11 executes the analysis process shown in FIG. 7 instead of the analysis process shown in FIG. Below, the details of the analysis processing executed by the processor 11 will be selectively described as a description of the second embodiment. It may be understood that the configuration of the information processing system 1 that is not mentioned in this embodiment is the same as in the first embodiment.

When starting the analysis process shown in FIG. 7, the processor 11 acquires the first data set 15A and the second data set 15B to be subjected to data fusion (S310, S320), as in the first embodiment.

The processor 11 generates a feature vector x for each first entity based on the first data set 15A (S310), similar to the process at S110. The processor 11 generates a feature vector y for each second entity based on the second data set 15B (S320), similar to the process at S120.

Furthermore, the processor 11 generates a group of low-dimensional feature vectors Dx corresponding to the group of feature vectors x, and a group of low-dimensional feature vectors Dx corresponding to the group of feature vectors y by the dimension reduction process, similarly to the processes in S130 and S140. A set of feature vectors Dy is generated (S330).

In subsequent S340, the processor 11 executes the same processes as those in S150, S160, and S170. That is, the processor 11 uses a group of low-dimensional feature vectors Dx to generate a similarity matrix K for the first set of entities, and uses a group of low-dimensional feature vectors Dy to generate a similarity matrix K for the second set of entities. Generate a similarity matrix L.

Further, the processor 11 uses the similarity matrix K and the similarity matrix L to search for the matrix Ω that maximizes the value Z(Ω) described in the first embodiment as the matrix Ω ^* (S340). Here, the searched matrix Ω ^* is expressed as a correspondence matrix Ω ^* .

After that, the processor 11 determines whether or not the repetition end condition is satisfied (S350). When determining that the repetition end condition is not satisfied (No in S350), the processor 11 executes the process of S360.

At S360, the processor 11 searches for a dimensionality reduction scheme that minimizes the cost of the Gromov-Wasserstein distance while fixing the correspondence matrix Ω ^* searched at S340.

Fixing the correspondence matrix Ω ^* corresponds to fixing the correspondence between the first entity and the second entity. Searching the matrix Ω that maximizes the value Z(Ω) as described above as the correspondence matrix Ω ^* is performed in the first feature space to fit the distribution of the second entity in the second feature space This corresponds to finding a map for mapping the plurality of first entities above to the second feature space.

The Gromov-Wasserstein distance cost corresponds to the transportation cost in the optimal transportation problem between the first entity and the second entity when mapping the first set of entities to the second feature space.

The cost of the Gromov-Wasserstein distance can be calculated using the similarity matrices K, L and the correspondence matrix Ω ^* . The similarity matrix K is, as described above, a matrix whose elements are the degrees of similarity between the first entities calculated based on the reduced-dimensional feature vector Dx. The similarity matrix L is a matrix whose elements are similarities between the second entities calculated based on the reduced-dimensional feature vector Dy.

Searching for a dimensionality reduction scheme that minimizes the cost of the Gromov-Wasserstein distance, the low dimensionality that best justifies the correspondence between the first entity and the second entity denoted by the correspondence matrix Ω ^* It corresponds to searching for a dimensionality reduction scheme for generating feature vectors Dx, Dy.

According to the correspondence matrix Ω ^* , the cost minimization is the distance in the feature space between the first entity and the second entity that correspond to each other, in other words, the low-dimensional feature vector of the first entity It corresponds to searching for a dimensionality reduction scheme that reduces the distance in the feature space between Dx and the low-dimensional feature vector Dy of the second entity.

For example, when transforming an M1-dimensional feature vector x into an M-dimensional low-dimensional feature vector Dx, the transform matrix Tx of M rows and M1 columns is applied to the feature vector x. When transforming an M2-dimensional feature vector y into an M-dimensional low-dimensional feature vector Dy, a transformation matrix Ty of M rows and M2 columns is applied to the feature vector y. At this time, the number of parameters m constituting the transformation matrices Tx and Ty is (M*M1+M*M2).

　The search for the dimension reduction method is realized by searching for the parameter m that minimizes the above-mentioned cost as the parameter m of the transformation matrixes Tx and Yy, for example, by using the gradient method or the like.

After that, the processor 11 reduces the dimension of the feature vectors x, y by the searched dimension reduction method (for example, transformation matrices Tx, Ty), and calculates new low-dimensional feature vectors Dx, Dy (S370).

The processor 11 uses the similarity matrix K based on the new low-dimensional feature vector Dx and the similarity matrix L based on the new low-dimensional feature vector Dy to create a matrix Ω that maximizes the value Z (Ω), Search as a new correspondence matrix Ω ^* (S340).

By repeatedly executing the processes of S360, S370, and S340 in this manner, the processor 11 re-searches the correspondence matrix Ω ^* with high matching accuracy along with a better dimensionality reduction method.

When the repetition end condition is satisfied (Yes in S350), the processor 11 executes the process of S380. The repetition end condition is satisfied, for example, when the process of S340 is executed a predetermined number of times, or when the amount of change in the correspondence matrix Ω ^* due to the re-search becomes less than a certain amount.

In S380, the processor 11 converts each of the first entities to at ^least correspond to one. Processor 11 is further capable of storing and outputting a correspondence table describing the correspondence between the first entity and the second entity.

Thereafter, the processor 11 performs data fusion processing in the same manner as in S190 to combine the first data set 15A and the second data set 15B to generate the extended data set 15C. , the generated extended data set 15C is stored in the storage 15 (S390).

The information processing system 1 of the second embodiment described above can associate the first entity and the second entity with even higher accuracy through the iterative process described above. Therefore, it is possible to generate the extended data set 15C with high precision.

<Third embodiment>
The information processing system 1 of the third embodiment is configured such that the processor 11 executes the evaluation process shown in FIG. Details of the evaluation process executed by the processor 11 will be described below as a description of the third embodiment. It may be understood that the configuration of the information processing system 1 that is not mentioned in this embodiment is the same as in the first or second embodiment.

The evaluation process is performed to evaluate whether or not the data set to be evaluated is an excellent data set capable of performing the matching and data fusion in the analysis process shown in FIG. 2 or 7 with high accuracy. be. The data set to be evaluated corresponds to a data set that can be used as the first data set 15A or the second data set 15B in the analytical process.

When starting the evaluation process, the processor 11 acquires the evaluation target data set specified by the user along with the execution instruction (S410). The processor 11 can acquire the designated evaluation target data set from the storage 15 .

After that, the processor 11 generates a first feature vector x_1 and a second feature vector x_2 for each entity based on the data set to be evaluated (S420). The data set to be evaluated may comprise, for each entity, feature data representing features of the corresponding entity with (Q1+Q2) elements.

The processor 11 can divide the (Q1+Q2) elements into a first element group consisting of Q1 elements and a second element group consisting of Q2 elements. Each of the (Q1+Q2) elements can be randomly classified into either the first element group or the second element group.

Based on the data set to be evaluated, the processor 11 generates, for each entity, a first feature vector x_1 describing a feature of the first element group of the corresponding entity and a feature of the second element group of the corresponding entity. and a second feature vector x_2 describing

For example, the data set to be evaluated is an element that can generate a feature vector v=(v[1], v[2], v[3], . When the number Q = (Q1 + Q2) of feature data is provided for each entity, a first feature vector x_1 containing Q1 elements = (v[1], v[2], ..., v[Q1]) and Q2 A second feature vector x_2=(v[Q1+1], v[Q1+2], . . . , v[Q1+Q2]) may be generated containing elements of

The first feature vector x_1 corresponds to the feature vector x for each entity in the first set of entities, and the second feature vector x_2 corresponds to the entity in the same second set of entities as the first set of entities. corresponding to each feature vector y.

After that, the processor 11 performs the same processing as the processing performed in S130 to S170 on the first feature vector x_1 and the second feature vector x_2 in S430 and S440.

In S430, the processor 11 performs dimension reduction processing on the first feature vector x_1 for each first entity and the second feature vector x_2 for each second entity, similarly to the processing in S130 and S140. , generate a low-dimensional feature vector Dx_1 and a low-dimensional feature vector Dx_2 having the same number of dimensions.

Based on the low-dimensional feature vector Dx_1 for each first entity, the processor 11 generates a similarity matrix representing the similarity of the low-dimensional feature vector Dx_1 between the first entities corresponding to the similarity matrix K. The processor 11 further generates a similarity matrix representing the similarity of the low-dimensional feature vectors Dx_2 between the second entities corresponding to the similarity matrix L based on the low-dimensional feature vectors Dx_2 for each second entity.

Based on these similarity matrices, the processor 11 searches for the matrix Ω that maximizes the value Z(Ω) as the correspondence matrix Ω ^* (S440).

After that, the processor 11 determines that the correspondence matrix Ω ^* for the first set of entities corresponding to the group of low-dimensional feature vectors Dx_1 and the second set of entities corresponding to the group of low-dimensional feature vectors Dx_2 is: A score is calculated to indicate the degree of correct representation of the correspondence between the first entity and the second entity (S450).

Thereby, the processor 11 evaluates whether or not the dataset to be evaluated is an excellent dataset capable of performing matching and data fusion by analysis processing with high accuracy (S450).

The processor 11 stores the correct correspondence relationship between the first entity and the second entity when generating the feature vector x_1 for each first entity and the feature vector x_2 for each second entity in advance in S420. can be kept.

Processor 11 calculates correspondence matrix Ω ^* by executing processing similar to the analysis processing in S430 and S440 in an environment in which the correct correspondence relationship is stored as described above, and calculates correspondence specified from correspondence matrix Ω ^* . Compare the relationship with the correct answer.

For example, the processor 11 performs the process of associating each of the first entities with one of the second entities based on the correspondence matrix Ω ^* in the same manner as in S180 and S380.

The processor 11 determines that the association is successful when the first entity and the second entity associated based on the correspondence matrix Ω ^* are the same entity in the data set to be evaluated. If they are not the same entity, it is determined that the association has failed.

The processor 11 can calculate the percentage of successful association among all entities as the score of the evaluation target data set (S450). After that, the processor 11 outputs the calculated score as an evaluation result (S460), and ends the evaluation process.

If matching and data fusion based on a single dataset cannot be performed with high accuracy, the dataset must contain sufficient information or data structure to achieve high accuracy matching and data fusion of the features of the set. It can be assumed that they do not.

This lack of information also affects the accuracy when performing analysis processing on two different data sets to perform matching and data fusion. Therefore, according to the evaluation process described above, it is possible to infer in advance whether the data set to be evaluated is a data set in which data fusion without a common variable can be executed with high accuracy.

In S460, the processor 11 can inform the user of the information processing system 1 whether or not the dataset to be evaluated is an excellent dataset by outputting the score. Thereby, the user can employ an appropriate combination of the first data set 15A and the second data set 15B for analysis processing to obtain the extended data set 15C with high reliability.

In order to obtain the desired extended data set 15C, an environment is conceivable in which it is sufficient to employ one of a plurality of mutually similar data sets as the first data set 15A to be combined with the second data set 15B.

For example, consider combining the first data set 15A regarding purchasing behavior and the second data set 15B regarding website visit behavior/web content browsing behavior to generate an extended data set 15C. In this case, it may be sufficient to generate the extended data set 15C using a data set relating to customer purchasing behavior of any one of the plurality of distribution organizations as the first data set 15A.

Examples of multiple distribution organizations include multiple convenience store chains. The data set on purchases of each convenience store chain may contain information on the same kind of purchasing behavior as that of other convenience store chains as consumer purchasing behavior.

Therefore, as the first data set 15A, it is considered sufficient to generate the extended data set 15C using a data set on customer purchasing behavior of any one of a plurality of convenience store chains.

In the above-described evaluation process, when there are multiple data sets as candidates for the first data set 15A (or second data set 15B), the accuracy of matching and data fusion is determined from these multiple data sets. can be used to select the optimal data set in terms of as the first data set 15A (or the second data set 15B).

For example, the processor 11 executes the selection process shown in FIG. One candidate can be adopted as a data set for data fusion. The data set targeted for data fusion in S110 and S310 corresponds to the first data set 15A, and the data set targeted for data fusion in S120 and S320 corresponds to the second data set 15B.

When the selection process shown in FIG. 9 is started, the processor 11 acquires multiple data sets as multiple data set candidates for data fusion (S510). The processor 11 can acquire a plurality of data sets designated by the user from the storage 15 .

After that, the processor 11 sets one of the plurality of datasets as the dataset to be evaluated (S520), and executes the evaluation process shown in FIG. 8 (S530). The processor 11 sets each data set as a data set to be evaluated (S520) and executes the evaluation process (S530) until the evaluation process for all of the plurality of data sets is executed (Yes in S540). repeat the process. As a result, the score calculated in S450 is obtained for each data set.

When evaluation processing is performed on all of the plurality of data sets and scores are obtained (Yes in S540), the processor 11 adopts the data set with the highest score among the plurality of data sets as the data set to be subjected to data fusion. (S550). After that, the selection process ends. At S110, S120, S310, S320, processor 11 may generate a feature vector (x or y) based on the adopted data fusion target dataset.

By executing the selection process in this way and selecting the optimum data set from a plurality of candidates, it is possible to generate the extended data set 15C with high accuracy.

Additionally, in the example of purchasing behavior, multiple dataset candidates for data fusion may include multiple datasets representing consumer purchasing behavior with different parameters. For example, the first candidate may be a data set capable of generating a feature vector whose elements include the number of items purchased for each consumer as an entity. A second candidate may be a data set capable of generating a feature vector whose elements include the purchase amount of each product for each consumer as an entity.

Preparing a plurality of data sets that explain similar features with different parameters and selecting a data set suitable for data fusion leads to the generation of a better extended data set 15C.

<Fourth embodiment>
The distribution system 30 of the fourth embodiment shown in FIG. 10 uses the data fusion technology of the first embodiment or the second embodiment to provide an external data set 35A, which is a data set provided from outside the distribution system 30, This system combines an internal data set 35B, which is a data set held inside the distribution system 30, and distributes advertisements based on an extended data set 35C generated thereby.

The distribution system 30 includes a processor 31, a memory 33, a storage 35, and a communication interface 39, as shown in FIG. Processor 31 executes processing according to computer program Pr1 stored in storage 35 . The storage 35 further comprises an internal data set 35B.

As shown in FIG. 11, the internal data set 35B includes, for each user, feature data describing the features of the corresponding user's online behavior in association with the corresponding user's advertisement ID. The advertisement ID, as is well known, is an identification code used for advertisement and is an ID unique to the information terminal.

The feature data associated with the advertisement ID describes the features of the user's online behavior observed through the information terminal assigned the corresponding advertisement ID. Online behavior includes website visit behavior and web content viewing behavior.

The distribution system 30 is connected to the wide area network through the communication interface 39 and provides an advertisement distribution service via the wide area network. A company-side system 40 that is a company-side system that uses the advertisement distribution service provides the distribution system 30 with distribution designation information together with advertisement content to be distributed. Advertising content is information content for advertising. The distribution designation information includes target designation information that designates distribution targets and distribution number designation information that designates the number of distributions.

The user company side system 40 further provides the delivery system 30 with a customer data set, which is a data set describing the characteristics of the customer corresponding to the delivery destination candidate, as an external data set 35A.

A customer data set can be, for example, a data set that describes the characteristics of the purchasing behavior of customers who use stores operated by the user company. For example, the customer data set may comprise, as feature data for each customer, feature data describing the purchase volume for each item of the corresponding customer regarding a plurality of items.

When a distribution request is input from the user company system 40 through the communication interface 39, the processor 31 executes the distribution control process shown in FIG. 12 based on the computer program Pr1.

When the distribution control process is started, the processor 31 receives from the user company system 40 the advertising content to be distributed, the distribution designation information including the target designation information and the distribution number designation information, and the customer data set as the external data set 35A. (S610).

After that, the processor 31 uses the external data set 35A as the first data set 15A and further uses the internal data set 35B as the second data set 15A to perform the same processing as in S110 to S190 in the analysis process. Execute the process. Thereby, processor 31 combines external data set 35A and internal data set 35B to generate extended data set 35C (S620).

By combining the external data set 35A and the internal data set 35B, the feature data for each customer contained in the external data set 35A includes the advertisement ID of the user who is highly likely to be the same person as the customer contained in the internal data set 35B. is associated.

The extended data set 35C comprises extended data in which the feature data of the corresponding customer's external data set 35A and the feature data of the corresponding user's internal data set 35B are combined for each entity. Each extension data is associated with the corresponding user's advertisement ID in the internal data set 35B.

　An entity here is a combination of a customer and a user that are associated with each other through data fusion. Data fusion creates a one-to-one correspondence between customers and users. For example, the extended data set 35C is a data set having a structure in which the advertisement ID of each entity is described in columns having "ID2_1", "ID2_2", and "ID2_3" illustrated in the extended data set 15C shown in FIG. obtain.

The processor 31 then calculates a score regarding the likelihood that each entity in the extended data set 35C is a delivery target (S630). For example, if the external data set 35A is a data set relating to customer purchasing behavior and the internal data set 35B is a data set relating to user online behavior, the processor 31 may determine the purchasing behavior characteristics of each entity in the extended data set 35C. The data and feature data about online behavior are input into a predetermined function to calculate a score that quantifies the likelihood that the corresponding entity is a distribution target.

A distribution target is a group of consumers who are targeted for distribution narrowed down by parameters that characterize consumers such as gender, age, purchasing tendency, online behavior tendency, interest, and interest, and is specified through target designation information.

After calculating the score in S630, the processor 31 selects the entities corresponding to the number of distributions specified by the user company system 40 in descending order of the calculated score among the group of entities associated with the advertisement ID, A content delivery destination is determined (S640). In this way, the processor 31 selects at least some of the plurality of users corresponding to the internal data set 35B associated with one of the plurality of customers corresponding to the external data set 35A as distribution destinations of the advertising content. .

After that, the processor 31 transmits the advertising content provided from the user company system 40 to the determined information terminal of the content delivery destination through the wide area network (S650). The advertisement content is distributed to the information terminal identified from the advertisement ID of the content distribution destination. After that, the processor 31 ends the distribution control process.

According to the distribution system 30 of the fourth embodiment described above, by combining the external data set 35A and the internal data set 35B using the data forsion technique without common variables, the customer whose advertisement ID is unknown The advertisement ID can be associated with the feature data of . As a result, it is possible to appropriately distribute the advertisement content to the customer of the external data set 35A whose advertisement ID is unknown.

<Fifth embodiment>
The distribution system 30 of the fifth embodiment is configured such that the processor 31 executes distribution control processing shown in FIG. 13 instead of the distribution control processing shown in FIG. Below, the details of the distribution control process executed by the processor 31 will be selectively described as a description of the fifth embodiment. It may be understood that the configuration of the distribution system 30 not mentioned in this embodiment is the same as in the fourth embodiment.

In this embodiment, when a distribution request is input from the user company system 40 through the communication interface 39, the processor 31 executes the distribution control process shown in FIG.

When the distribution control process starts, the processor 31 acquires the advertising content to be distributed, the distribution designation information, and the customer data set as the external data set 35A from the user company system 40 (S710).

The distribution designation information acquired in S710 does not include target designation information, but only distribution number designation information. The customer data set acquired as the external data set 35A is a specific customer data set that describes the characteristics of the customer group corresponding to the distribution target narrowed down by the user company.

After that, the processor 31 combines the external data set 35A and the internal data set 35B to generate the extended data set 35C (S720), similar to the processing in S620. The extended data set 35C includes extended data obtained by combining the feature data of the corresponding customer's external data set 35A and the feature data of the corresponding user's internal data set 35B for each entity.

In the processing of S720 of the present embodiment, there is no result that the customer of the external data set 35A is associated with all the users of the internal data set 35B. The extended data set 35C of the present embodiment also includes, as extended data of one entity, user feature data that is not associated with the customer of the company using the data. This extended data is the feature data of the corresponding user that the internal data set 35B has, which is not substantially extended.

In this embodiment, among the group of entities corresponding to the extended data set 35C, the group of entities associated with the group of customers corresponding to the external data set 35A is referred to as a seed, and the other group of entities is referred to as a seed. is expressed as non-seed.

After the process of S720, the processor 31 calculates the similarity of the feature indicated by the internal data set 35B between each non-seed entity and each seed entity based on the extended data set 35C (S730). The similarity can be calculated by the feature space distance between each non-seed entity and each seed entity.

After calculating the degree of similarity, the processor 31 determines, as distribution destinations, the number of entities corresponding to the number of distributions specified by the distribution designation information in descending order of similarity (S740). At this time, all entities corresponding to the seed are also determined as delivery destinations.

In this way, the processor 31 selects a set of seeds, which is a set of users associated with a plurality of customers corresponding to the external data set 35A, and seeds and features among a plurality of users corresponding to the internal data set 35B. is selected as a distribution destination of advertising content.

After that, similar to the process of S650, the processor 31 transmits the advertising content provided from the user company system 40 to the information terminal of the content delivery destination determined in S740 through the wide area network (S750). After that, the distribution control process is terminated.

According to the distribution system 30 of this embodiment described above, based on the data set of the customer group provided from the user company side system 40, information terminals of a larger group of consumers exhibiting similar characteristics to the customer group can deliver advertising content. Therefore, according to this embodiment, it is possible to efficiently distribute advertisements to many consumers.

<Sixth embodiment>
The distribution system 30 of the sixth embodiment is configured to provide a prediction service along with an advertisement distribution service similar to the distribution system 30 of the fourth or fifth embodiment.

In this embodiment, the processor 31 executes the prediction process shown in FIG. 14 in response to an execution request from the user company system 40. Details of the prediction process executed by the processor 31 will be selectively described below as a description of the sixth embodiment. It may be understood that the configuration of the distribution system 30 not mentioned in this embodiment is the same as in the fourth or fifth embodiment.

When starting the prediction process, the processor 31 acquires the data set to be analyzed from the user company system 40 through the communication interface 39 together with the analysis condition designation information (S810). The data set to be analyzed is a data set comprising feature data for each customer to be analyzed.

The analysis condition specifying information can be information specifying the target product for which the customer's purchase possibility is to be evaluated. In the prediction process, the possibility that each analysis target customer will purchase a designated target product is predicted by calculating a predicted value of the number of purchases of the target product. Prediction here corresponds to estimating the customer's behavior, and predicted value corresponds to an estimate of the behavior.

After executing the process of S810, the processor 31 uses the data set to be analyzed as the first data set 15A and further uses the internal data set 35B as the second data set 15B to perform S110 to S170 in the analysis process. Alternatively, by executing the same processing as the processing of S310 to S370, a correspondence matrix Ω ^* indicating the correspondence between each customer to be analyzed and each user having feature data in the internal data set 35B is calculated (S820). .

Further, the processor 31 extracts a predetermined number of users close to the corresponding customer for each customer to be analyzed based on the calculated correspondence matrix Ω ^* , and extracts the above extracted users who can be identified from the internal data set 35B. Based on the weighted average of the number of product purchases, the predicted number of purchases of the target product by the corresponding customer is calculated (S830). Thus, the processor 31 infers the customer's purchasing behavior from the associated user's purchasing behavior. The internal data set 35B includes information that can identify the number of purchases of target products by each user.

Each element of the correspondence matrix Ω ^* indicates the degree of similarity between the customer and the user with a value of 0-1. The element in the i-th row and j-th column of the correspondence matrix Ω ^* is the i-th user in the set of users corresponding to the internal data set 35B and the j-th user in the set of customers corresponding to the data set to be analyzed. The similarity between the customers of and is indicated by a value of 0 to 1.

A weighted average is calculated, for example, using similarity as a weight. A weighted average can be calculated as follows, assuming that the first, second, and third users are extracted as three users close to the customer.

That is, the degree of similarity between the customer and the first user is w1, the degree of similarity between the customer and the second user is w2, and the degree of similarity between the customer and the third user is w3. is p1, the number of purchases of the target product by the second user is p2, and the number of purchases of the target product by the third user is p3, the predicted value pe of the number of purchases of the target product by the customer is pe=(w1 *p1+w2*p2+w3*p3)/3.

From the correspondence matrix Ω ^* , it is possible to identify the degree of similarity (in other words, magnitude of correspondence) with all users for each customer. Therefore, without the process of extracting close users, a weighted average of the number of purchases of the target product by all users may be used to calculate the expected number of purchases of the target product by the customer.

After executing the process of S830, the processor 31 outputs prediction data describing the predicted number of purchases of corresponding products for each customer to the source of the prediction process execution request (S840). After that, the processor 31 terminates the prediction process shown in FIG.

According to another example, after executing the process of S830, the processor 31, instead of or in addition to outputting the prediction data, uses data in descending order of the prediction value based on the prediction value of the number of purchases of corresponding products for each customer. A process of distributing the advertising content promoting the purchase of the target product to the number of customers corresponding to the number of distributions specified by the company may be executed (S840).

The distribution system 30 of the sixth embodiment has been described above, but according to this embodiment, a meaningful advertisement distribution service can be provided using data fusion technology without common variables, and furthermore, a meaningful marketing solution can be provided. can be provided.

[others]
It goes without saying that the present disclosure is not limited to the embodiments described above, and can take various forms. A function possessed by one component in the above embodiment may be distributed to a plurality of components. Functions possessed by multiple components may be integrated into one component. A part of the configuration of the above embodiment may be omitted. At least part of the configurations of the above embodiments may be added or replaced with respect to the configurations of other above embodiments. All aspects included in the technical ideas specified by the language in the claims are embodiments of the present disclosure.

Claims (15)

1. a first acquisition unit configured to acquire a first data set relating to a plurality of first entities, the first data set describing characteristics of each of the plurality of first entities;
   a second acquisition unit configured to acquire a second data set relating to a plurality of second entities, the second data set describing characteristics of each of the plurality of second entities;
   A group of first feature vectors identified from the first data set, each of the first feature vectors representing features of a corresponding one of the plurality of first entities. A group of first feature vectors and a group of second feature vectors identified from the second data set, each of the second feature vectors being one of the plurality of second entities. a set of first low-dimensional feature vectors corresponding to the set of first feature vectors by performing a dimensionality reduction operation on a set of second feature vectors representing features of a corresponding entity of; generating a group of second low-dimensional feature vectors corresponding to the group of the second feature vectors and having the same number of dimensions as the group of the first low-dimensional feature vectors; a dimensionality reduction unit configured to:
   Each of the plurality of first entities is associated with at least one of the plurality of second entities based on the group of first low-dimensional feature vectors and the group of second low-dimensional feature vectors. a mapping unit configured to:
   An information processing system comprising
2. The associating unit determines the degree of similarity between the first entities identified from the group of the first low-dimensional feature vectors, and the second entity identified from the group of the second low-dimensional feature vectors. each of the plurality of first entities to the plurality of 2. The information processing system according to claim 1, associated with at least one of the second entities.
3. The first set of low-dimensional feature vectors is defined by a first feature space,
   The second set of low-dimensional feature vectors is defined by a second feature space,
   The associating unit identifies, from the group of the second low-dimensional feature vectors, the distribution of the plurality of first entities in the first feature space identified from the group of the first low-dimensional feature vectors. for mapping the plurality of first entities on the first feature space to the second feature space to match the distribution of the plurality of second entities in the second feature space where 3. The information processing system according to claim 1 or 2, wherein a mapping of is searched for, and each of said plurality of first entities is associated with at least one of said plurality of second entities based on said mapping.
4. The associating unit is a formula comprising a matrix K, a matrix L, and a matrix H
   

   

   
   A matrix Ω that maximizes the value Z(Ω) according to is searched as a matrix Ω ^* , and based on said matrix Ω ^* , each of said plurality of first entities is replaced by at least one of said plurality of second entities configured to map to
   the number of said first entities is N;
   the number of said second entities is the same as said first entity;
   In the matrix K, the value of the i-th row and j-th column element represents the similarity between the i-th entity and the j-th entity among the plurality of first entities, and a first low-dimensional feature vector of the i-th entity among the entities of and a first low-dimensional feature vector of the j-th entity among the plurality of first entities is a first similarity matrix of N rows and N columns,
   In the matrix L, the value of the i-th row and j-th column element represents the similarity between the i-th entity and the j-th entity among the plurality of second entities, and the plurality of second entities and the second low-dimensional feature vector of the i-th entity among the entities of and the second low-dimensional feature vector of the j-th entity among the plurality of second entities, is a second similarity matrix of N rows and N columns,
   The matrix H is a matrix of N rows and N columns in which the value of the element in the i-th row and j-th column indicates a value of 1−1/N when i=j, and indicates a value of 0 when i≠j. The information processing system according to claim 1 or 2.
5. The associating unit improves the matrix Ω ^* by repeatedly performing re-search processing on the matrix Ω ^* until a predetermined condition is satisfied, and based on the improved matrix Ω ^* , the plurality of configured to associate each of the first entities with at least one of the plurality of second entities;
   The re-search process is
   changing a dimensionality reduction method in the dimensionality reduction process based on the matrix Ω ^* ;
   causing the dimensionality reduction unit to execute the dimensionality reduction process based on the changed dimensionality reduction method;
   re-searching the matrix Ω ^* based on the thereby newly obtained group of first low-dimensional feature vectors and group of second low-dimensional feature vectors;
   5. The information processing system according to claim 4, comprising:
6. The associating unit selects the dimensionality reduction method from the group of the first low-dimensional feature vectors and the group of the second low-dimensional feature vectors. 6. The information processing system according to claim 5, wherein the change is made so as to shorten the distance in the feature space to the low-dimensional feature vector.
7. the first data set includes a plurality of first feature data, each of the plurality of first feature data representing a feature of a corresponding one of the plurality of first entities;
   the second data set includes a plurality of second feature data, each of the plurality of second feature data representing a feature of a corresponding one of the plurality of second entities;
   The information processing system is
   Based on the association between the plurality of first entities and the plurality of second entities by the associating unit, the plurality of second feature data are added to each of the plurality of first feature data. an extended data set comprising a plurality of extended data by combining one of them, each of said plurality of extended data is a combination of a corresponding one of the first feature data and the second feature data The information processing system according to any one of claims 1 to 6, further comprising: a data fusion unit that generates an extended data set containing a plurality of extended data that are data.
8. the first entity and the second entity are people;
   the first data set is a data set describing a first characteristic of each of a plurality of people belonging to a first group;
   The information processing system according to any one of claims 1 to 7, wherein said second data set is a data set describing second characteristics of each of a plurality of people belonging to a second group.
9. The combination of the first feature and the second feature includes a feature related to purchasing behavior, a feature related to movement in at least one of an online space and an offline space, and/or a plurality of points on the space. 10. The information processing system according to claim 8, which is a combination of: a feature relating to a visit;
10. the first entity and the second entity are people;
    the second data set is associated with identification information of an information terminal corresponding to each of the plurality of second entities;
    The information processing system is
    Selection of selecting at least a part of a set of the plurality of second entities that are associated with any one of the plurality of first entities by the association unit as a delivery destination of the information content Department and
    a distribution unit configured to distribute the information content to a group of information terminals corresponding to distribution destinations of the information content based on the identification information;
    The information processing system according to any one of claims 1 to 9, comprising:
11. The selecting unit selects a first set, which is a set of the second entities associated with any one of the plurality of first entities by the associating unit, and the plurality of second entities, the 11. The information processing system according to claim 10, wherein a second set similar in characteristics to the first set is selected as a distribution destination of the information content.
12. the first entity and the second entity are people;
    the second data set describes behavioral characteristics of each of the plurality of second entities;
    The information processing system further includes:
    an estimating unit that calculates, for each entity of interest, an estimated value regarding the behavior of the corresponding entity of interest with respect to one or more entities of interest;
    the one or more entities of interest are at least part of the plurality of first entities;
    12. The information processing according to any one of claims 1 to 11, wherein said estimated value is calculated based on a feature relating to behavior of at least one of said plurality of second entities associated with said corresponding entity of interest. system.
13. A computer-implemented information processing method comprising:
    obtaining a first data set for a plurality of first entities, the first data set describing characteristics of each of the plurality of first entities;
    obtaining a second data set for a plurality of second entities, the second data set describing characteristics of each of the plurality of second entities;
    A group of first feature vectors identified from the first data set, each of the first feature vectors representing features of a corresponding one of the plurality of first entities. A group of first feature vectors and a group of second feature vectors identified from the second data set, each of the second feature vectors being one of the plurality of second entities. a set of first low-dimensional feature vectors corresponding to the set of first feature vectors by performing a dimensionality reduction operation on a set of second feature vectors representing features of a corresponding entity of; generating a group of second low-dimensional feature vectors corresponding to the group of the second feature vectors and having the same number of dimensions as the group of the first low-dimensional feature vectors; and
    Each of the plurality of first entities is associated with at least one of the plurality of second entities based on the group of first low-dimensional feature vectors and the group of second low-dimensional feature vectors. and
    Information processing method including.
14. said associating
    the similarity between the first entities identified from the group of the first low-dimensional feature vectors and the similarity between the second entities identified from the group of the second low-dimensional feature vectors each of the plurality of first entities to one of the plurality of second entities such that the correlation between the first entities in terms of similarity matches the correlation between the second entities based on 14. The information processing method according to claim 13, comprising:
15. A computer-readable recording medium storing a computer program containing instructions for causing a computer to execute the information processing method according to claim 13 or claim 14.

Images