JP6079270B2 - Information provision device - Google Patents

Description

  The present invention relates to an information providing apparatus, and more particularly to an information providing apparatus that provides a topic that activates communication between users. The present invention also relates to an information providing program and an information providing method.

  Communication with various communication tools is performed in a virtual space, such as sending diary information on social networking sites and blogs, commenting on them, and e-mails. And in order to promote such communication, there exists a system which recommends the topic used as the subject of communication.

  The topic that is the subject of communication is considered to be a topic that is preferred by the user to be recommended and the communication partner. As a technique for recommending such a topic, there is collaborative filtering.

  Non-Patent Document 1 describes a collaborative filtering method. First, an evaluation history is input. The evaluation history is history information that evaluates a topic that each user likes. The evaluation includes, for example, evaluation by a user in five stages, and evaluation in two stages based on whether or not the user has read based on log information indicating that a topic has been read. Next, a user who is close to the recommendation target user (hereinafter referred to as a preference-similar user) is searched. The closeness of preference is calculated by the similarity of the user column vector in the matrix expression of the evaluation history. Here, the matrix expression of the evaluation history is a matrix expression in which a row is a topic, a column is a user, and a value is an evaluation value. That is, the user column vector is a vector having each element as an evaluation value for each topic. The similarity is calculated using cosine similarity, Pearson correlation coefficient, or the like. Next, a topic that has been highly appreciated by a preference-similar user and that has not yet been browsed by the recommendation target user is recommended to the recommendation target user.

Hijikata Yoshinori, “1. Preference Extraction and Information Recommendation Technology (Basic Theory of Preference Extraction and Information Recommendation, <Special Feature> Utilizing User's Preference-The Forefront of Preference Extraction Technology”), Information Processing 48 (9), 957-965, 2007 Thomas Hofmann, “Probabilistic Latent Semantic Indexing,” Proceedings of the Twenty-Second Annual International SIGIR Conference on Research and Development in Information Retrieval (SIGIR-99), 1999 David M. Blei, Andrew Y. Ng, Michael I. Jordan, “Latent Dirichletallocation,” Journal of Machine Learning Research 3 (4-5), pp. 993-1022. T. L. Griffiths et al., “Finding scientific topics,” Proc. Of the National Academy of Sciences of the United States of America, vol.101, pp.5228-5235, 2004 Koide et al., “Document Clustering Using Relationships Between Words as Constraints”, Proc. 8 (2), 269-270 Andrzejewski et al., “Incorporating domain knowledge into topic modeling via DirichletForest priors”, Proceedings of the 26th International Conference on MachineLearning (ICML2009), 25-32

  However, the above-described technique causes the following problems. First, since it cannot be determined whether the topic is the same type of interest area, a topic that is interested in a preference-similar user but not interested in a recommendation target user has been recommended. For this reason, it is not possible to recommend a topic that is preferred by both a recommendation target user and a preference-similar user as a communication subject, and thus it has not become a subject that activates communication.

  Human interest is divided into several areas of interest. Therefore, even if a region of interest is the same for a recommendation target user and a preference-similar user, another region of interest is not necessarily the same. For example, it is assumed that there are two users who give a favorable evaluation to a topic related to politics, user 1 and user 2. When the user 2 gives a good evaluation to a topic related to performing arts, the user 1 does not necessarily like the topic of performing arts.

  As described above, collaborative filtering does not distinguish the regions of interest and recommends even topics of different regions of interest. For this reason, there has been a problem in that it is not possible to provide a topic that is a subject of communication that is recommended by both the recommendation target user and the preference similar user.

  For this reason, an object of the present invention is to provide an information providing apparatus, a program, and an information providing method capable of solving the above-described problem that a topic that is of interest to a user cannot be provided. There is.

An information providing apparatus according to one aspect of the present invention
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Dimensional compression means for compressing a single dimension and outputting a compressed dimension vector representation of a topic and a compressed dimension vector representation of a user;
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
With
The configuration is as follows.

Moreover, the program which is the other form of this invention is:
In the information processing device,
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Dimensional compression means for compressing a single dimension and outputting a compressed dimension vector representation of a topic and a compressed dimension vector representation of a user;
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
It is a program for realizing.

An information providing method according to another aspect of the present invention includes:
Enter evaluation history information consisting of evaluations for each topic for each user,
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Are compressed into one dimension, and the compressed dimension vector expression of the topic and the compressed dimension vector expression of the user are output,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Mapping,
Based on the group, for each user, extract a group to which the user belongs, extract a topic belonging to the group, select a recommended topic candidate based on the element value of the compression dimension vector representation of the topic,
Among the recommended topic candidates, providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information,
The configuration is as follows.

  By being configured as described above, the present invention can provide a topic that is of interest to the user.

It is a block diagram which shows the structure of the whole system containing the topic recommendation apparatus in Embodiment 1 of this invention. It is a flowchart explaining the flow of the whole operation | movement of Embodiment 1 of this invention. It is a figure which shows the example of the evaluation log | history storage means shown in FIG. It is a figure which shows the example of the two-dimensional matrix which the dimension compression means shown in FIG. 1 produces. It is a figure which shows the example of the title of a topic of the evaluation log | history storage means shown in FIG. It is a conceptual diagram explaining the compression method of the dimension compression means shown in FIG. It is a figure which shows the example of an output of the grouping means shown in FIG. It is a figure which shows the example of an output of the recommendation topic candidate creation means shown in FIG. It is a block diagram which shows the structure of the whole system containing the topic recommendation apparatus in Embodiment 2 of this invention. It is a flowchart explaining the flow of the whole operation | movement of Embodiment 2 of this invention. It is a figure which shows the example of the communication other party storage means shown in FIG. It is a block diagram which shows the structure of the whole system containing the topic recommendation apparatus of Embodiment 3 of this invention. It is a flowchart explaining the flow of the whole operation | movement of Embodiment 3 of this invention. It is a figure which shows the example of the user model storage means shown in FIG. It is a figure which shows the example of the content model storage means shown in FIG. It is a block diagram which shows the structure of the whole system containing the topic recommendation apparatus of Embodiment 4 of this invention. It is a flowchart explaining the flow of the whole operation | movement of Embodiment 4 of this invention. It is a block diagram which shows an example of the computer which implement | achieves the topic recommendation apparatus in Embodiment 1-3 of this invention.

<Embodiment 1>
[Device configuration]
Hereinafter, a topic recommendation device, a topic recommendation method, and a program according to Embodiment 1 of the present invention will be described with reference to FIGS. 1 to 8 and FIG.

  First, the configuration of the topic recommendation device (information providing device) in the first embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a topic recommendation device according to Embodiment 1 of the present invention.

  As shown in FIG. 1, the topic recommendation device according to the first exemplary embodiment includes an evaluation history storage unit 121, an input unit 101, a dimension compression unit 102, a grouping unit 103, a recommendation candidate creation unit 104, and a recommendation. Means 105. Each of these units generally operates as follows.

  The evaluation history storage unit 121 stores evaluation history information representing a history of evaluation of each user on each topic. Specifically, a user, a topic, and an evaluation are included. In addition, there may be an evaluation time.

  The input unit 101 receives input of evaluation history information from the evaluation history storage unit 121 and passes it to the dimension compression unit 102.

  The dimension compression unit 102 converts the input evaluation history information into a topic compression vector representation and a user compression vector representation. Specifically, first, a two-dimensional matrix is created from the evaluation history information, with topics as rows, users as columns, and evaluation values as values. Then, the two-dimensional matrix is compressed with a user who has evaluated the same topic and a topic evaluated by the same user as one dimension. Next, the compressed dimension vector representation of the topic having the compressed dimension as an element and the compressed dimension vector representation of the user are output.

  A method for compressing a two-dimensional matrix is calculated by a topic model. As a specific calculation method, Probabilistic latent semantic analysis (PLSA) shown in Non-Patent Document 2 and Latent Dirichlet allocation (LDA) shown in Non-Patent Document 3 are used. In the second embodiment to be described later, a restriction is imposed so that communication partners are compressed into one dimension.

  The topic model is a model that expresses a document with a distribution of K latent topics based on a hypothesis that a document group consisting of a plurality of words is input and a document is generated from K latent topics. A latent topic is a category that potentially exists in a document, and the distribution of the latent topic is expressed by the distribution of a word group. With regard to the word distribution of latent topics, a word group that tends to appear in the same document has a high probability. As a result, the document is represented by a K-dimensional vector whose elements are K potential topics and whose values are the probabilities of the latent topics. A word is expressed as a K-dimensional vector with elements as K potential topics and values as probabilities of potential topics.

  In the present embodiment, the dimension is compressed by the topic model with the document as the topic, the word as the user, and the word frequency as the evaluation value. That is, the word distribution of the latent topic is represented by the user group distribution. As a result, users who have evaluated the same document have a high probability in the same potential topic. A topic is expressed as a K-dimensional vector with elements as K potential topics and values as probabilities of potential topics. The user expresses an element as a K-dimensional vector having K potential topics and a value as the probability of the latent topic.

  The grouping means 103 receives an input of a topic compression dimension vector and a user compression dimension vector, and a topic having a compression dimension value equal to or greater than a first threshold in the topic compression dimension vector representation for each dimension. And users having a compression dimension value greater than or equal to the second threshold among the user's compression dimension vectors are grouped together as one group.

  The recommended topic candidate creation unit 104 receives the input of the group, extracts a plurality of groups to which the user belongs for each user, extracts a topic group corresponding to the group, and selects the topic in the order of the scores of the compression dimension. Line up and create recommended topic candidates.

  The recommendation unit 105 receives the input of the evaluation history information and the recommended topic candidate, and recommends a topic that the user has not yet evaluated for each user to be recommended. At this time, among the recommended topic candidates arranged in order by the recommended topic candidate creation unit 104, the number of topics corresponding to the number of recommended topics set in advance from the top is recommended.

[Device operation]
Next, the operation of the topic recommendation device according to the first embodiment will be described in detail with reference to FIGS. FIG. 2 is a flowchart for explaining the overall operation flow of the first embodiment. In the following description, FIG. 1 is taken into consideration as appropriate. In the first embodiment, the topic recommendation method is implemented by operating the topic recommendation device. Therefore, the description of the topic recommendation method in the first embodiment is replaced with the following description of the operation of the topic recommendation device.

[Matrix representation of evaluation history]
First, the input means 101 receives evaluation history information from the evaluation history storage means 121, and the dimension compression means 102 converts the evaluation history information into a matrix representation (step S1).

  An example of the evaluation history information stored in the evaluation history storage unit 121 is shown in FIG. The evaluation history information includes a user ID, a topic ID, an evaluation value, and an evaluation time. Here, the evaluation value may be a result of the user explicitly assigning 5 levels or the like, or may be 1 or the like from a log in which the user browses the content implicitly. In FIG. 3, assuming that the evaluation is implicit, the evaluation value is 1 when the user browses the content. Further, “...” In the figure indicates omission (the same applies hereinafter).

  The dimension compression means converts the evaluation history information into a two-dimensional matrix having a topic as a row, a user as a column, and an evaluation value as a value. An example of matrix representation is shown in FIG. FIG. 4 represents the topic ID, user ID, and evaluation value of the evaluation history information as a row, a column, and a value, respectively. For example, it is indicated that the user 1 has evaluated the topic 1 with the evaluation value 1. An evaluation value of 0 indicates that no evaluation was made.

  An example of a topic corresponding to the topic ID is shown in FIG. Topics 1 and 2 are topics related to politics, topics 3 and 4 are topics related to politics and performing arts, and topics 5 and 6 are topics related to performing arts. Compared to the matrix representation example, user 1 and user 2 evaluate topics related to political interest areas, user 3 and user 4 evaluate topics related to two different areas of interest political and entertainment, It can be seen that 5 and user 6 are evaluating topics related to the entertainment related interest areas.

[Dimension compression of evaluation history]
Next, the dimension compression means 102 dimensionally compresses the matrix representation (step S2). A topic model method is used for dimension compression. Below, the technique by PLSA is shown.

First, PLSA represents the simultaneous probability of topic x and user y using latent topic z as follows.
P (x, y) = P (x) Σ_ {z ∈ Z} P (y | z) Pr (x | z)
The joint probability P (x, y) indicates that the user y has evaluated the content x.

  The parameters of the PLSA model can be solved by an EM algorithm or the like. Using the results, P (z | x) and P (z | y) are obtained, and the probability that topic x is represented by K latent topics and user y are represented by K latent topic models, respectively. Probability. In other words, the compression vector representation of the topic is represented by the probability distribution of the topic latent topic, specifically, the element is the latent topic and the value is P (z = j | x = d) (kth of the dth topic (Probability of potential topic). The user's compressed vector representation is expressed by the probability distribution of the user's latent topic, the element is the latent topic, the value is P (z = k | y = u) (the probability of the kth latent topic of the uth user) and The vector is

  The topic modeling is done in the same way for the LDA model. The solution of the LDA model is described in Non-Patent Document 4 above.

  Non-Patent Document 4 introduces a method based on a Gibbs sampler as follows. As described above, topic modeling by LDA is performed using a topic as a document and a user as a word.

First, a topic is randomly assigned to each user. Next, r is sampled randomly from the uniform distribution. Next, for each appearance position of the user, the value of the following formula 1 is calculated. If the value of Equation 1 is greater than or equal to r, the topic number is incremented. Otherwise, the topic is assigned to the user's appearance position. Finally, P (z | x) and P (z | y) are obtained from the assigned topic numbers.

The above i is the appearance position of the user, j is the topic number assigned to the appearance position of the user, and K is the number of topics. The function Q calculates the following equation (2).

Here, the function q is expressed as the following equation (3).

Here, the value of the equation 4 in the equation 3 represents the frequency of the topic number j other than the position i among the user numbers u at the position i.

The value of equation (5) in the above equation (3) represents the frequency that the topic number is j at a place other than the position i regardless of the user number.

The value of Equation 6 in Equation 3 above represents the frequency at which the topic number is j at a place other than position i among the topic numbers d at position i.

The value of Equation 7 in Equation 3 represents the frequency other than the position i in the topic number d.
W is the number of users, K is the number of topics, and α and β are prior distribution parameters.

  FIG. 6 shows a conceptual diagram of generation of a latent topic in the topic model. The topic model assumes a model in which a topic has several potential topics and the potential topics are evaluated by the user. Accordingly, the topics with the same evaluation user have the same probability distribution of the topic latent topics. In addition, users with the same evaluated topic have the same probability distribution of the user's latent topics. From this, Topic 1 and Topic 2, Topic 3 and Topic 4, Topic 5 and Topic 6 each have a probability distribution P (z | d) of similar latent topics. Further, user 1 and user 2, user 3 and user 4, and user 5 and user 6 have similar potential topic probability distributions P (z | y), respectively. In order to obtain a probability distribution in which these are similar, a portion surrounded by a dotted line in FIG. 6, that is, a topic evaluated by the same user, a user's portion evaluated by the same topic becomes a latent topic, and how much is this potential topic portion Probability distributions of potential topics are created depending on whether they are included.

When the number of latent topics K is 2, it is assumed that a probability distribution of latent topics is generated as follows.
Compressed vector representation of topic 1 = P (z | x = topic 1) = (0.8,0.2)
Compressed vector representation of topic 2 = P (z | x = topic 2) = (0.7,0.3)
Compression vector representation of topic 3 = P (z | x = topic 3) = (0.5, 0.5)
Compressed vector representation of topic 4 = P (z | x = topic 4) = (0.5, 0.5)
Compression vector representation of topic 5 = P (z | x = topic 5) = (0.3,0.7)
Compression vector representation of topic 6 = P (z | x = topic 6) = (0.2,0.8)
Compressed vector representation of user 1 = P (z | y = user 1) = (0.7,0.3)
Compressed vector representation of user 2 = P (z | y = user 2) = (0.8,0.2)
Compressed vector representation of user 3 = P (z | y = user 3) = (0.5,0.5)
Compressed vector representation of user 4 = P (z | y = user 4) = (0.5,0.5)
Compressed vector representation of user 5 = P (z | y = user 5) = (0.2,0.8)
Compressed vector representation of user 6 = P (z | y = user 6) = (0.3,0.7)

  Here, the number of latent topics K may be determined in advance. Alternatively, the total number of regions of interest may be estimated from the number of regions of interest of one user, the average number of communication partners of one user, and the total number of users, and this value may be used as the number of potential topics. Here, the total number of users can be obtained by referring to the evaluation history storage unit 121.

Total number of regions of interest = number of different regions of interest / number of users of one region of interest Number of different regions of interest = number of regions of interest of one user x total number of users Number of users of one region of interest = average number of communication partners + 1

For example, if the number of interest areas of one user is 2, the number of communication partners of one user is 3, and the total number of users is 6, the total number of interest areas is calculated as follows. Let the total number of areas of interest be the number of potential topics.
Total number of areas of interest = 2 × 6 ÷ (3 + 1) = 3

  By estimating in this way, the number of potential topics can be automatically adjusted according to an increase in users. Further, even if the number of users increases, it is effective because the number of interest areas of one user and the communication partner of one user are considered not to change so much.

[Group]
Next, the grouping means 103 creates a group with the compressed vector expression of the topic and the compressed vector expression of the user as input (step S3). The user's compressed vector expression and the topic's compressed vector expression are referred to, and for each vector element, the user and topic having a value equal to or greater than a specific threshold value are grouped together.

  For the above-described user compressed vector representation and topic compressed vector representation, the user threshold is set to 0.5 and the topic threshold is set to 0.5. Referring to the first element of the user's compressed vector expression and the topic's compressed vector expression, the users above the threshold are User 1, User 2, User 3, User 4, and the topics above the threshold are Topic 1, Topic 2, Topic 3 and Topic 4. Put this together as a group. The second element is similarly grouped together.

  An example of the generated group is shown in FIG. You can combine groups, users, and topics. The user indicates a plurality of users separated by commas, and indicates the probability value of the potential topic by separating “:”. The topics also indicate a plurality of topics separated by commas, and the probability values of latent topics are separated by “:”.

  In FIG. 7, user 3 and user 4 belong to both group 1 and group 2. Topic 3 and Topic 4 also belong to both Group 1 and Group 2. Referring to FIGS. 4 and 5, the user 3 and the user 4 are interested in topics related to politics and performing arts, the group 1 represents politics and the group 2 represents entertainment related. Therefore, user 3, user 4, topic 3, and topic 4 belong to both groups.

  In this example, the user threshold and the topic threshold may be determined in advance. Alternatively, the user threshold may be determined from the number of regions of interest of one user and the user compressed vector expression, and the topic threshold may be determined from the user threshold, evaluation history information, and topic compressed vector expression. It is determined as follows. This is determined by searching the user's threshold value in half until the average number of groups to which one user belongs becomes the same as the number of interest areas of one user. Group users into groups based on user thresholds. Next, referring to the evaluation history information, the topics evaluated by the users in each group are acquired. Next, referring to the compressed vector expression of the topic, the probability value of the latent topic of the topic is acquired, and the lowest value is set as a threshold value.

[Nomination candidate creation]
Next, the recommended topic candidate creation means 104 selects a topic candidate to be recommended for each user and creates a recommended topic candidate (step S4).

  The recommended topic candidate creation unit 104 receives input of a correspondence table of groups, users, and topics from the grouping unit 103, and extracts a group to which each user belongs. Next, a topic corresponding to the group is extracted. When the topic is extracted from a plurality of groups, the total is used as a score. Next, the topics are arranged in the order of scores, and the candidate candidates are recommended in order from the highest score.

Here, the topic score can be expressed by a mathematical formula as follows.
Topic i score = Σ_ {k ∈ G} P (z = k | d = i)
(G represents the set of groups to which topic i belongs.)

  FIG. 8 shows an example of topic candidates to be recommended for each user created by the recommended topic candidate creation unit 104. This is an example in which recommended topic candidates for user 1, user 2, and user 3 are created using the correspondence table of groups, users, and topics in FIG. 7 as input. Create groups and topic candidates for each user. Here, the user 3 belongs to two groups, and the topic 3 and the topic 4 belong to the group 1 and the group 2. Therefore, the scores of Topic 3 and Topic 4 are calculated as follows using the above mathematical formula.

Topic 3 score = P (z = Group 1 | d = Topic 3) + P (z = Group 2 | d = Topic 3)
= 0.5 + 0.5 = 1.0
Topic 4 score = P (z = Group 1 | d = Topic 4) + P (z = Group 2 | d = Topic 4)
= 0.5 + 0.5 = 1.0

[Recommended topic decision]
Next, the recommendation means 105 receives a recommended topic candidate as input, refers to the evaluation history information, and recommends a topic to the user (step S5).

  The recommendation unit 105 refers to the evaluation history information, determines a topic that has not yet been evaluated by the user among the recommended topic candidates, and outputs it as a recommendation target topic. For example, for the user 1, the topic 2 is set as the recommendation target topic.

[program]
The program according to the first embodiment of the present invention may be a program that causes a computer (information processing apparatus) to execute the processing from step S1 to step S5 shown in FIG. Or what is necessary is just a program which makes a computer implement | achieve each means shown to the codes | symbols 101-105 shown in FIG. By installing and executing this program on a computer, the topic recommendation device and the topic recommendation method according to the first embodiment can be realized.

  In the case of Embodiment 1 of the present invention, the CPU (Central Processing Unit) of the computer functions as the input means 101, the dimension compression means 102, the grouping means 103, the recommended candidate topic creation means 104, and the recommendation means 105, Perform processing. In the first embodiment, the evaluation history accumulating unit 121 stores the data files constituting these in a storage device such as a hard disk provided in the computer, or a recording medium on which the data files are stored. Is implemented in a reading device connected to a computer.

  Here, a computer that realizes the topic recommendation device by executing the program according to the first embodiment will be described with reference to FIG. FIG. 18 is a block diagram illustrating an example of a computer that implements the topic recommendation device according to Embodiment 1 of the present invention.

  As shown in FIG. 18, the computer 150 includes a CPU 111, a main memory 112, a storage device 113, an input interface 114, a display controller 115, a data reader / writer 116, and a communication interface 117. These units are connected to each other via a bus 121 so that data communication is possible.

  The CPU 111 performs various operations by expanding the program (code) in the first embodiment stored in the storage device 113 in the main memory 112 and executing them in a predetermined order. The main memory 112 is typically a volatile storage device such as a DRAM (Dynamic Random Access Memory). The program according to the first embodiment is provided in a state where it is stored in a computer-readable recording medium 120. Note that the program in the present embodiment may be distributed on the Internet connected via the communication interface 117.

  Specific examples of the storage device 113 include a hard disk and a semiconductor storage device such as a flash memory. The input interface 114 mediates data transmission between the CPU 111 and an input device 118 such as a keyboard and a mouse. The display controller 115 is connected to the display device 119 and controls display on the display device 119. The data reader / writer 116 mediates data transmission between the CPU 111 and the recording medium 120, and reads a program from the recording medium 120 and writes a processing result in the computer 110 to the recording medium 120. The communication interface 117 mediates data transmission between the CPU 111 and another computer.

  Specific examples of the recording medium 120 include general-purpose semiconductor storage devices such as CF (Compact Flash) and SD (Secure Digital), magnetic storage media such as a flexible disk, or CD-ROM (Compact Disk). Optical storage media such as Read Only Memory).

[effect]
In this way, the topic evaluated by the same user from the evaluation history and the user who evaluated the same topic are compressed to the same dimension, and the user group and the topic group are grouped based on the compressed dimension. It is possible to create a group with a dimension as an area of interest, and to extract a topic to be recommended for each group to which the user belongs, thereby performing topic recommendation for each area of interest.

<Embodiment 2>
[Device configuration]
Next, a second embodiment of the present invention will be described. FIG. 9 is a block diagram illustrating a configuration of the entire system including the topic recommendation device according to the second embodiment. The second embodiment is different from the first embodiment shown in FIG. 1 in that a communication partner storage unit 211 is provided. Furthermore, the operation of the dimension compression means 102 is different. Below, a different point is mainly demonstrated.

  The communication partner storage unit 221 stores the user and the communication partner. This data may be created in advance. In addition, based on the information in the evaluation history storage unit 121, users who are similar in the evaluated topic may be added as communication partners. Since users who are similar to the evaluated topics are users who are likely to become new communication partners by providing a common topic, these users may be added.

  The dimension compression means 102 receives the evaluation history information as input, refers to the communication partner storage means 221, and converts it into a compressed vector expression of the topic and a compressed vector expression of the user. Specifically, first, a two-dimensional matrix is created from the evaluation history information, with the topic as the row, the user as the column, and the evaluation value as the value. Next, the communication partner storage unit 221 is referred to, and the two-dimensional matrix is compressed using the users of the communication partners and the topics evaluated by the same user as one dimension after evaluating the same topic. Next, the compressed dimension vector representation of the topic having the compressed dimension as an element and the compressed dimension vector representation of the user are output.

[Device operation]
Next, the operation of the topic recommendation device of the second embodiment will be described in detail with reference to the flowchart of FIG. FIG. 10 is a flowchart for explaining the overall operation flow of the second embodiment.

  As shown in FIG. 10, the second embodiment is different from the first embodiment in that a step of reading a communication partner (step S21) is added. Further, the operation of the step of compressing the matrix (step S2) is different. The following description will mainly focus on the different steps.

  First, as in the first embodiment, the input unit 101 receives the evaluation history information, and the dimension compression unit 102 converts the evaluation history information into a matrix representation (step S1).

[Read communication partner]
Next, the dimension compression unit 102 refers to the communication partner storage unit 221 and reads communication partner information (step S21). The communication partner storage unit 221 stores user relationships between communication partners. An example is shown in FIG. User 1 knows that user 2 and user 3 are communication partners. The communication partner information is read for each pair such as (user 1, user 2) and (user 1, user 3).

[Compress matrix]
Next, as in the first embodiment, the two-dimensional matrix is dimensionally compressed (step S2). However, at the time of dimension compression, the communication partner information is referred to, and calculation is performed with a constraint that increases the probability value that the communication partners are the same topic.

  Referring to the technique of Non-Patent Document 5 above, which is a technique for restricting conditions between words in PLSA, a constraint condition between users is used. Below, the content of the said nonpatent literature 6 is demonstrated first, and it expands to the constraint conditions between users next.

In Non-Patent Document 5, the probability distribution of topic P (z) is changed. Based on the relationship between words, if the change is related to a word but assigned to another potential topic, a penalty term for adding a penalty is added.
P (z) = Π P (z_i) × exp {-Σ _ {<i, j> ∈C} δ (z_i ≠ z_j) × w_ij} ÷ G

  Here, z_i is the latent topic of the word at the i-th position, C is a constraint set indicating the relationship between the i-th word and the j-th word, and δ (•) is the latent potential of the i-th word A function that becomes 1 when the topic and the potential topic of the j-th word are different, w_ij is a penalty weight (> 0), and G is a normalization term.

  In the present embodiment, the relationship between words is applied to the relationship between users, and when the users are communication partners, this is set as a constraint set C. If the communication partner is assigned to another potential topic, a penalty is imposed. That is, in this example, when (User 1, User 2) and (User 1, User 3) are assigned to different topics, a penalty is imposed.

  Similarly, in LDA, topic modeling may be performed by using the method of Non-Patent Document 6 in which a constraint condition between words is introduced. Even in Non-Patent Document 6, a case where there is a synonymous relationship between words is set as a constraint. Similarly, in this embodiment, communication partner information is added as a constraint condition.

  Further, as described below, the communication partner may be included as a restriction by changing the initial topic assignment and the function q of Non-Patent Document 4 described above.

  First, communication partners are set as a network, and the network is clustered by the number of topics T. Existing methods such as Newman method are used for network clustering. Assign a cluster number to each user as a topic. Next, r is sampled randomly from the uniform distribution. Next, the value of the following equation 8 is calculated for each appearance position of the user. If the value of equation (8) is greater than or equal to r, the topic number is incremented. Otherwise, the topic is assigned to the user's appearance position.

Here, the function q of Non-Patent Document 4 is replaced with a function to which the similarity between users is added as follows.

  Here, the function sim represents the similarity between two users. fu represents the communication partner user of user u. Fu represents the number of communication partner users of user u. The function sim, which is the similarity between users, uses a cosine similarity or the like using a topic viewed by the user as a vector, a cosine similarity using a user's communication partner number as a vector, or the like.

  In this way, topic assignment is performed by replacing the initial topic assignment and the function q with the communication partner as a constraint.

[Step S3 to Step S5]
Next, as shown in FIG. 10, steps S3 to S5 are executed. However, steps S3 to S5 shown in FIG. 10 are the same steps as steps S3 to S5 shown in FIG.

[program]
The program according to the second embodiment of the present invention may be a program that causes a computer to execute steps S1 to S5 shown in FIG. Or what is necessary is just a program which makes each computer implement | achieve each means shown to the codes | symbols 101-105 shown in FIG. By installing and executing this program on a computer, the topic recommendation device and the topic recommendation method in the second embodiment can be realized.

  In the case of Embodiment 2 of the present invention, the CPU (Central Processing Unit) of the computer functions as the input means 101, the dimension compression means 102, the grouping means 103, the recommended candidate topic creation means 104, and the recommendation means 105, Perform processing. In the second embodiment, the evaluation history storage unit 121 and the communication partner storage unit 221 store the data files constituting them in a storage device such as a hard disk provided in the computer, or this data file. Is stored in a reading device connected to a computer.

  Here, a computer that realizes the topic recommendation device by executing the program according to the second embodiment will be described with reference to FIG. FIG. 18 is a block diagram illustrating an example of a computer that implements the topic recommendation device according to Embodiment 2 of the present invention.

  As shown in FIG. 18, the computer 150 includes a CPU 111, a main memory 112, a storage device 113, an input interface 114, a display controller 115, a data reader / writer 116, and a communication interface 117. These units are connected to each other via a bus 121 so that data communication is possible.

  The CPU 111 performs various calculations by developing the program (code) in the second embodiment stored in the storage device 113 in the main memory 112 and executing them in a predetermined order. The main memory 112 is typically a volatile storage device such as a DRAM (Dynamic Random Access Memory). The program according to the second embodiment is provided in a state where it is stored in a computer-readable recording medium 120. Note that the program in the present embodiment may be distributed on the Internet connected via the communication interface 117.

  Specific examples of the storage device 113 include a hard disk and a semiconductor storage device such as a flash memory. The input interface 114 mediates data transmission between the CPU 111 and an input device 118 such as a keyboard and a mouse. The display controller 115 is connected to the display device 119 and controls display on the display device 119. The data reader / writer 116 mediates data transmission between the CPU 111 and the recording medium 120, and reads a program from the recording medium 120 and writes a processing result in the computer 110 to the recording medium 120. The communication interface 117 mediates data transmission between the CPU 111 and another computer.

  Specific examples of the recording medium 120 include general-purpose semiconductor storage devices such as CF (Compact Flash) and SD (Secure Digital), magnetic storage media such as a flexible disk, or CD-ROM (Compact Disk). Optical storage media such as Read Only Memory).

[effect]
As described above, when the dimension is compressed, the communication partner information is made a constraint condition so that the communication partners can easily belong to the same group. Thereby, when there is a communication partner in advance, it is possible to present a topic that interests the communication partner and to support communication.

<Embodiment 3>
Next, a third embodiment of the present invention will be described. FIG. 12 is a block diagram illustrating a configuration of the entire system including the topic recommendation device according to the third embodiment. The third embodiment is different from the second embodiment shown in FIG. 9 in that an evaluation information complementing unit 301, a user model storage unit 322, and a content model storage unit 321 are provided. Below, a different point is mainly demonstrated.

  The user model storage unit 322 stores a user model including an interesting keyword and its weight for each user. This information may acquire an interest keyword from a user beforehand, and may produce | generate from evaluation log | history information and the content model mentioned later.

  The content model storage unit 321 stores, for each topic, a content model composed of keywords existing in the topic and their weights. This information is generated by dividing a document described in a topic into words by morphological analysis or the like and using the words as keywords.

  The evaluation information complementing unit 301 receives the input of the evaluation history information, refers to the user model storage unit 322 and the content model storage unit 321, and refers to a topic not evaluated by the user, that is, a topic whose evaluation value is 0. For a pair with a user who has not been evaluated, the similarity is calculated from the user model and the content model. If the similarity is equal to or higher than a specific threshold, the topic evaluation value by the user is added to the evaluation history information.

[Device operation]
Next, the operation of the topic recommendation device of the third embodiment will be described in detail with reference to the flowchart of FIG. FIG. 13 is a flowchart for explaining the overall operation flow of the third embodiment.

  As shown in FIG. 13, the third embodiment is different from the second embodiment in that a step for supplementing the evaluation history information is added. The following description will mainly focus on the different steps.

[Estimation of evaluation information]
The evaluation information complementing unit 301 refers to the user model storage unit 322 and the content model storage unit 321 and adds evaluation information for a user / topic pair with no evaluation history (step S31).

  First, referring to the evaluation history, a user and topic pair with no evaluation history are extracted. Here, it is assumed that there is no evaluation history of user 1 and topic 2 (see FIG. 4). Next, the user model storage unit 322 and the content model storage unit 321 are referred to, and the user model of the user 1 and the content model of the topic 2 are acquired.

  An example of the user model storage unit 322 is shown in FIG. The user model storage unit 322 stores the user ID and its user model. A user model is a set of keywords and their weights. The keyword is a keyword that the user is interested in. For example, the user 1 is interested in “Prime Minister” and “Economy”. These keywords can also be obtained by learning with reference to evaluation history information and a content model described later. For example, there are a method of averaging a content model browsed by a user as a user model, a method of averaging by weighting with evaluation values, and the like.

  An example of the content model storage unit 321 is shown in FIG. The content model storage unit 321 stores the topic ID and its content model. A content model is a set of keywords and their weights. A keyword can be obtained from a word that appears in a topic, meta information of the topic, or the like.

Next, the similarity between the user model of user 1 and the content model of topic 2 is calculated. As the similarity, a method based on the Euclidean distance, an inner product, a cosine similarity, or the like is used. When inner product is used, it is calculated as follows.
SIM (user 1, topic 2) = 0 x 0.5 + 0.3 x 0.3 + 0.1 x 0.1 +…
Each term is a product related to “Prime Minister”, “Finance”, “Plan” and so on.

  Next, when the similarity exceeds a threshold value, an evaluation value is assigned. For example, when the threshold value 0.5 is exceeded, the similarity is assigned as the evaluation value.

[Step S1 to Step S5]
Next, as shown in FIG. 13, steps S1 to S5 are executed. However, steps S1 to S5 shown in FIG. 13 are the same steps as steps S1 to S5 shown in FIG. 2 or FIG.

[program]
The program according to the third embodiment of the present invention may be a program that causes a computer to execute steps S31 to S5 shown in FIG. Or what is necessary is just a program which makes each computer implement | achieve each means shown to the code | symbol 101-105,301 shown in FIG. By installing and executing this program on a computer, the topic recommendation device and the topic recommendation method according to the third embodiment can be realized.

  In the case of Embodiment 3 of the present invention, the CPU (Central Processing Unit) of the computer includes an input means 101, an evaluation information storage means 301, a dimension compression means 102, a grouping means 103, a recommended candidate topic creation means 104, and a recommendation Functions as means 105 and performs processing. In the third embodiment, the evaluation history storage unit 121, the communication partner storage unit 221, the user model storage unit 322, and the content model storage unit 321 are configured in a storage device such as a hard disk provided in the computer. This is realized by storing a data file or by mounting a recording medium storing the data file on a reading device connected to a computer.

  A computer that implements the topic recommendation device according to Embodiment 3 of the present invention has the same configuration as the computer that implements the topic recommendation device according to Embodiment 1 shown in FIG.

[effect]
When the evaluation history is very sparse, the dimension compression method using the topic model cannot accurately estimate the potential topic. Therefore, the region of interest cannot be specified with high accuracy, and a topic of interest between the user and the communication partner cannot be provided. By adopting such a configuration, it is possible to accurately identify the region of interest by predicting and complementing what kind of evaluation is performed on a topic that has not yet been evaluated by the user. Thereby, the topic in which the user and the communication partner are interested can be provided.

<Embodiment 4>
Next, a fourth embodiment of the present invention will be described. FIG. 16 is a block diagram showing a configuration of the entire system including the topic recommendation device according to the fourth embodiment. In the fourth embodiment, unlike the third embodiment shown in FIG. 12, the topic recommendation device is further provided with an evaluation history selecting means 401. The fourth embodiment is different from the third embodiment in terms of the function of the recommended topic candidate creation unit 104. Hereinafter, the different points will be mainly described.

  The evaluation history selection means 401 receives input of evaluation history information, specifies a topic with a small number of evaluation users or a user with a small number of evaluation topics, and selects evaluation history other than the topic and the user. Specifically, an unpopular topic that is not evaluated by many users is selected based on the evaluation history information. In addition, an unrated user who has not evaluated many topics is selected. Next, one or both of the unpopular topic and the unrated user are set, and other than these, that is, an evaluation history other than the unpopular topic and an evaluation history other than the unrated user are selected.

  The recommended topic candidate creation unit 104 extracts a plurality of groups to which the user belongs for each user based on the correspondence relationship between the group, which is an output of the grouping unit 103, the user, and the topic. Then, one or more topics corresponding to the group are extracted, and the score of each topic is determined from the score of the compression dimension of the topic and the similarity between the content model of each topic and the user model of the user. Arrange in order of score and create recommended candidate topics.

[Device operation]
Next, the operation of the topic recommendation device of the fourth embodiment will be described in detail with reference to the flowchart of FIG. FIG. 17 is a flowchart for explaining the overall operation flow of the fourth embodiment.

  As shown in FIG. 17, the fourth embodiment is different from the third embodiment in that step S41 for selecting evaluation information is added. The operation of step S4 for creating a recommendation candidate topic is also different. The following description will mainly focus on the different steps.

  First, as described in the third embodiment, the topic evaluation information is supplemented.

[Select Evaluation Information]
Next, the evaluation information selection means 401 selects evaluation history information based on the evaluation history and the supplemented evaluation information (step S41).

  First, the topic with few evaluation users is specified by inputting the evaluation history information. Specifically, a topic that is evaluated only from the number of users equal to or less than a specific threshold is specified as an unpopular topic. Furthermore, a user with a small number of evaluation topics is specified. Specifically, a user who has evaluated only the number of topics equal to or less than a specific threshold is specified as an unevaluated user.

  Next, either or both of the unpopular topic and the unevaluated user are set, and the evaluation history other than one or both, that is, the evaluation history other than the unpopular topic and the evaluation history other than the unevaluated user select.

[Step S1 to Step S3]
Subsequently, as in the third embodiment, the data is converted into a matrix representation, the dimensions are compressed, and grouping is performed.

[Step S4]
Next, the recommended topic candidate creation unit 104 receives the input of the correspondence relationship between the group, the user, and the topic, refers to the user model storage unit 322 and the content model storage unit 321, and creates a topic candidate recommended for each user. To do. In that case, the score of each topic is calculated from the value of each compression dimension and the similarity between the user model and the content model. Specifically, it calculates as follows about all the topics i.
Topic i score = w × Σ_ {k ∈ G} P (z = k | d = i) + (1-w) × SIM (user j, topic i)
(G represents the set of groups to which topic i belongs.)

  Here, w is a weight indicating whether or not to place importance on the probability value of the latent topic, and is in the range of 0 to 1. SIM (user j, topic i) indicates the similarity between user j and topic i, and is calculated from the user model and content model using the Euclidean distance, inner product, and cosine similarity.

[Step S5]
Subsequently, a recommended topic is identified and recommended as in the third embodiment.

[program]
The program according to the fourth embodiment of the present invention may be a program that causes a computer to execute steps S31 to S5 shown in FIG. Or what is necessary is just a program which makes a computer implement | achieve each means shown to the code | symbol 101-105,301,401 shown in FIG. By installing and executing this program on a computer, the topic recommendation device and the topic recommendation method according to the fourth embodiment can be realized.

In the case of the fourth embodiment of the present invention, the CPU (Central Processing Unit) of the computer is the input means 101, the evaluation information storage means 301, the evaluation information selection means 401, the dimension compression means 102, the grouping means 103, and the recommended candidate topic creation. Functions as the means 104 and the recommendation means 105 to perform processing. In the fourth embodiment, the evaluation history storage unit 121, the communication partner storage unit 221, the user model storage unit 322, and the content model storage unit 321 are configured in a storage device such as a hard disk provided in the computer. This is realized by storing a data file or by mounting a recording medium storing the data file on a reading device connected to a computer.
A computer that implements the topic recommendation device according to Embodiment 4 of the present invention has the same configuration as the computer that implements the topic recommendation device according to Embodiment 1 shown in FIG.

[effect]
When the evaluation history is very sparse, the dimension compression method using the topic model cannot accurately estimate the potential topic. Therefore, the region of interest cannot be specified with high accuracy, and a topic of interest between the user and the communication partner cannot be provided. By adopting such a configuration, even if there are new topics with a low number of evaluations or new users with a low number of evaluations, by creating a matrix that excludes these, the region of interest can be accurately defined. Can be identified. Thereby, the topic in which the user and the communication partner are interested can be provided.

  In addition, when a recommendation candidate is created, a topic score is calculated not only from a compressed dimension value but also from a user model and a content model, so that it can be recommended to excluded topics and users. Thereby, the evaluation number can be gradually increased.

<Appendix>
Part or all of the above-described embodiment can be described as in the following supplementary notes. The outline of the configuration of the information providing apparatus, the program, and the information providing method in the present invention will be described below. However, the present invention is not limited to the following configuration.

(Appendix 1)
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Dimensional compression means for compressing a single dimension and outputting a compressed dimension vector representation of a topic and a compressed dimension vector representation of a user;
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
An information providing apparatus comprising:

(Appendix 2)
An information providing apparatus according to attachment 1, wherein
Based on the evaluation history information, the dimension compression means uses latent topic modeling as an element of a potential topic that is a category that potentially exists in the topic, and the probability of the potential topic to which the topic belongs and the potential topic to which the user belongs. And calculating the topic's compression dimension vector representation and the user's compression dimension vector representation as a topic potential topic probability and a user potential topic probability,
The grouping means associates a topic having a potential topic probability equal to or higher than a first threshold and a user having a potential topic probability equal to or higher than a second threshold as one group for each potential topic,
The recommended topic candidate creation means extracts a group to which the user belongs for each user, extracts a topic belonging to the group, and selects a recommended topic candidate based on the probability of a potential topic of the topic.
Information providing device.

(Appendix 3)
The information providing apparatus according to appendix 1 or 2,
Based on the evaluation history information and communication partner information indicating the communication relationship of each user, the dimension compressing means combines users of communication partners who have evaluated the same topic and topics evaluated by the same user. Compress to one dimension and output the topic's compressed dimension vector representation and the user's compressed dimension vector representation,
Information providing device.

(Appendix 4)
An information providing device according to attachment 3, wherein
Based on the evaluation history information and the communication partner information, the dimension compression unit adds a latent topic that is a category potentially existing in a topic by latent topic modeling that weights communication partners as conditions. The probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs are calculated, and the compression dimension vector representation of the topic and the compression dimension vector representation of the user are represented by the probability of the topic potential topic and the user potential of the topic. Expressed as the probability of the topic,
Information providing device.

(Appendix 5)
An information providing apparatus according to any one of appendices 1 to 4,
The similarity between the user and the topic that is not evaluated by the user based on the user model that the user likes and the weight, the content model that includes the keyword and weight in the topic, and the evaluation history information And further comprising an evaluation information complementing means for complementing the evaluation value for the topic in the evaluation history information according to the similarity.
Information processing device.

(Appendix 6)
An information providing apparatus according to any one of appendices 1 to 5,
Based on the evaluation history information, the user to be evaluated is identified as having a small number of topics based on a preset criterion, and the number of topics evaluated based on a preset criterion is identified as the topic and / or Or further comprising evaluation history selection means for selecting evaluation information other than the user as a target for creating a two-dimensional matrix by the dimension compression means,
Information processing device.

(Appendix 7)
An information providing apparatus according to any one of appendices 1 to 6,
The recommended topic candidate creating means calculates the similarity between the topic and the user based on the user model that the user likes and weights, and the content model that includes the keywords and weights in the topic. From the value of the compressed dimension expression and the similarity between the topic and the user, calculate a score for each topic, and select a recommended topic candidate based on the score,
Information providing device.

(Appendix 8)
In the information processing device,
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Dimensional compression means for compressing a single dimension and outputting a compressed dimension vector representation of a topic and a compressed dimension vector representation of a user;
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
A program to realize

(Appendix 9)
Enter evaluation history information consisting of evaluations for each topic for each user,
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic Are compressed into one dimension, and the compressed dimension vector expression of the topic and the compressed dimension vector expression of the user are output,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Mapping,
Based on the group, for each user, extract a group to which the user belongs, extract a topic belonging to the group, select a recommended topic candidate based on the element value of the compression dimension vector representation of the topic,
Among the recommended topic candidates, providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information,
Information provision method.

(Appendix 10)
The information providing method according to appendix 9, wherein
Based on the evaluation history information, the potential topic modeling is a latent topic that is a category potentially existing in the topic, and calculates the probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs, The topic's compression dimension vector representation and the user's compression dimension vector representation are expressed as the topic's potential topic probability and the user's potential topic probability,
A topic having a potential topic probability equal to or higher than the first threshold and a user having a potential topic probability equal to or higher than the second threshold are associated as one group for each potential topic,
Extracting a group to which the user belongs for each user, extracting a topic belonging to the group, and selecting a recommended topic candidate based on the probability of a potential topic of the topic.
Information provision method.

  Note that the above-described program is stored in a storage device or recorded on a computer-readable recording medium. For example, the recording medium is a portable medium such as a flexible disk, an optical disk, a magneto-optical disk, and a semiconductor memory.

  Although the present invention has been described with reference to the above-described embodiment and the like, the present invention is not limited to the above-described embodiment. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  As described above, according to the present invention, it is possible to provide a topic that is interesting to both the user and the communication partner in consideration of the field of interest, so that communication between users can be activated. Therefore, the present invention can be applied to an application such as an information providing support device in a communication tool or a program for realizing the information providing support device on a computer.

101 input means 102 dimension compression means 103 grouping means 104 recommended topic candidate creation means 105 recommendation means 121 evaluation history storage means 221 communication partner storage means 301 evaluation information supplement means 321 content model storage means 322 user model storage means 401 evaluation information selection means 111 CPU
112 Main memory 113 Storage device 114 Input interface 115 Display controller 116 Data reader / writer 117 Communication interface 118 Input device 119 Display device 120 Recording medium 121 Bus 150 Computer

Claims (9)

An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic , The probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs by using latent topic modeling as a potential topic. calculated, and the compression-dimensional vector representation of a topic represented by the probability of the topic belongs potential topics, and dimensionality reduction means for outputting the compression-dimensional vector representation of the user represented by the probability of latent topic to which the user belongs,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
An information providing apparatus comprising: An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information and communication partner information indicating the communication relationship of each user , create a two-dimensional matrix with topics as rows, users as columns, and user evaluation values of topics as values Potentially exists in the topic by latent topic modeling that weights the communication partner as a condition when compressing the user of the communication partner who evaluated the topic and the topic evaluated by the same user into one dimension Calculate the probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs, using the potential topic that is the category as an element, and the compressed dimension vector representation of the topic expressed by the probability of the potential topic to which the topic belongs, and the latency to which the user belongs dimensions for outputting a compressed-dimensional vector representation of the user represented by the probability of topics And contraction means,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
An information providing apparatus comprising: The information providing device according to claim 1 or 2 ,
The similarity between the user and the topic that is not evaluated by the user based on the user model that the user likes and the weight, the content model that includes the keyword and weight in the topic, and the evaluation history information And further comprising an evaluation information complementing means for complementing the evaluation value for the topic in the evaluation history information according to the similarity.
Information processing device. The information providing device according to any one of claims 1 to 3 ,
Based on the evaluation history information, the user who evaluates identifies a topic with a small number based on a preset criterion and a user with a small number of topics evaluated based on a preset criterion, and the topic And / or further comprising evaluation history selection means for selecting evaluation information other than the user as a target for creating a two-dimensional matrix by the dimension compression means,
Information processing device. The information providing device according to any one of claims 1 to 4 ,
The recommended topic candidate creating means calculates the similarity between the topic and the user based on the user model that the user likes and weights, and the content model that includes the keywords and weights in the topic. From the value of the compressed dimension expression and the similarity between the topic and the user, calculate a score for each topic, and select a recommended topic candidate based on the score,
Information providing device. In the information processing device,
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic , The probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs by using latent topic modeling as a potential topic. calculated, and the compression-dimensional vector representation of a topic represented by the probability of the topic belongs potential topics, and dimensionality reduction means for outputting the compression-dimensional vector representation of the user represented by the probability of latent topic to which the user belongs,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
A program to realize In the information processing device,
An input means for inputting evaluation history information comprising evaluations for each topic of each user;
Based on the evaluation history information and communication partner information indicating the communication relationship of each user , create a two-dimensional matrix with topics as rows, users as columns, and user evaluation values of topics as values Potentially exists in the topic by latent topic modeling that weights the communication partner as a condition when compressing the user of the communication partner who evaluated the topic and the topic evaluated by the same user into one dimension Calculate the probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs, using the potential topic that is the category as an element, and the compressed dimension vector representation of the topic expressed by the probability of the potential topic to which the topic belongs, and the latency to which the user belongs dimensions for outputting a compressed-dimensional vector representation of the user represented by the probability of topics And contraction means,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Grouping means to associate,
A recommended topic candidate that extracts a group to which the user belongs for each user based on the group, extracts a topic that belongs to the group, and selects a recommended topic candidate based on an element value of a compression dimension vector representation of the topic Creating means;
Among the recommended topic candidates, topic providing means for providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information;
A program to realize Enter evaluation history information consisting of evaluations for each topic for each user,
Based on the evaluation history information, a two-dimensional matrix having a topic as a row, a user as a column, and a topic evaluation value by the user as a value is created, and a topic evaluated by the same user as a user who evaluated the same topic , The probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs by using latent topic modeling as a potential topic. calculated, and outputs the compressed-dimensional vector representation of a topic represented by the probability of the topic belongs potential topics, and compression-dimensional vector representation of the user represented by the probability of potential topics to which the user belongs, a,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Mapping,
Based on the group, for each user, extract a group to which the user belongs, extract a topic belonging to the group, select a recommended topic candidate based on the element value of the compression dimension vector representation of the topic,
Among the recommended topic candidates, providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information,
Information provision method. Enter evaluation history information consisting of evaluations for each topic for each user,
Based on the evaluation history information and communication partner information indicating the communication relationship of each user , create a two-dimensional matrix with topics as rows, users as columns, and user evaluation values of topics as values Potentially exists in the topic by latent topic modeling that weights the communication partner as a condition when compressing the user of the communication partner who evaluated the topic and the topic evaluated by the same user into one dimension Calculate the probability of the potential topic to which the topic belongs and the probability of the potential topic to which the user belongs, using the potential topic that is the category as an element, and the compressed dimension vector representation of the topic expressed by the probability of the potential topic to which the topic belongs, and the latency to which the user belongs outputs, and compression-dimensional vector representation of the user represented by the probability of the topic,
The topic whose element value of the compression dimension vector expression of the topic is equal to or more than the first threshold and the user whose element value of the compression dimension vector expression of the user is equal to or more than the second threshold are grouped into one group for each vector element. Mapping,
Based on the group, for each user, extract a group to which the user belongs, extract a topic belonging to the group, select a recommended topic candidate based on the element value of the compression dimension vector representation of the topic,
Among the recommended topic candidates, providing the user with a topic that has not yet been evaluated by the user who is the subject of the topic based on the evaluation history information,
Information provision method.