Program, information processing apparatus, and information processing method

Abstract

To optimize the processing in recommending a document to a user.SOLUTION: A program causes a computer to execute an acquisition procedure, a probability calculation procedure, a similarity calculation procedure, and an extraction procedure. The acquisition procedure acquires an affiliation probability that first document data belong to each of document clusters according to browsing trend of document data, the document cluster being formed by inputting document data to a first clustering model. The probability calculation procedure calculates an affiliation probability that second document data browsed by a user belongs to each of the document clusters, using a prediction model generated on the basis of the first clustering model. The similarity calculation procedure calculates similarities between the affiliation probability for each of the document clusters calculated for the first document data, and the affiliation probability for each of the document clusters calculated for the second document data. The extraction procedure extracts document data to be recommended to the user out of the first document data, on the basis of the similarity.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to a program, an information processing device, and an information processing method. [Background technology]

[0002] Conventionally, techniques for providing content suitable for a user have been proposed. For example, Patent Document 1 discloses a method for determining user-specific information that may be of interest to a user in the field of providing recommended information to the user. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2012-512465 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the above-mentioned conventional techniques have room for improvement in terms of optimizing the process of recommending documents to a user.

[0005] For example, the above-mentioned conventional technologies merely identify specific terms that are sufficiently likely to be of interest to the user in order to assist in providing recommended information to the user, but do not take into consideration further optimizing the process for providing recommended information.

[0006] Therefore, the present invention proposes a program, an information processing device, and an information processing method that are capable of optimizing the process of recommending documents to a user. [Means for solving the problem]

[0007] In order to solve the above problem, one embodiment of a program according to the present invention causes a computer to execute the following steps: an acquisition step for acquiring a probability that first document data belongs to each document cluster according to a viewing tendency of the document data, the probability calculation step for calculating a probability that second document data viewed by a user belongs to each of the document clusters using a predictive model generated based on the first clustering model; a similarity calculation step for calculating a similarity between the probability of belonging for each document cluster calculated for the first document data and the probability of belonging for each document cluster calculated for the second document data; and an extraction step for extracting document data to be recommended to the user from the first document data based on the similarity. Effect of the Invention

[0008] According to the present invention, it is possible to optimize the process of recommending documents to a user. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of a system according to an embodiment. [Diagram 2] FIG. 2 is a diagram showing an overall view of information processing according to the embodiment. [Diagram 3] FIG. 3 is an explanatory diagram for explaining explanatory variables in the learning phase in detail. [Figure 4] FIG. 4 is a diagram illustrating an example of the configuration of a learning device according to an embodiment. [Diagram 5] FIG. 5 is a diagram illustrating an example of generation of learning data according to the embodiment. [Figure 6] FIG. 6 is a diagram illustrating an example of the configuration of a recommendation server according to the embodiment. [Figure 7] FIG. 7 is a flowchart illustrating the operation procedure of the learning device. [Figure 8] FIG. 8 is a flowchart illustrating the operation procedure of the recommendation server 100. As shown in FIG. [Figure 9] FIG. 9 is a block diagram illustrating an example of a hardware configuration of an information processing device according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings. In this specification and the drawings, components having substantially the same functional configurations are denoted by the same reference numerals, and duplicated explanations will be omitted.

[0011] One or more embodiments (including examples, modifications, and application examples) described below can be implemented independently. Meanwhile, at least a part of the embodiments described below may be implemented in appropriate combination with at least a part of another embodiment. These embodiments may include novel features different from each other. Thus, these embodiments may contribute to solving different purposes or problems and may provide different effects.

[0012] Furthermore, since the term "document" generally refers to a medium (e.g., an electronic file) on which a sentence is written, it is understood to encompass the concept of "article content." For this reason, in the following embodiments, "document" will be replaced with "article" for explanation. In this case, for example, "document data" can be written as "article data." "Article data" is composed of sentences indicating the title of the article and sentences indicating the main text of the article.

[0013] (Embodiment) 1. Introduction Conventionally, there is a method to determine the article content to be recommended to a user based on the "similarity" of the article content. However, there is a problem that the viewpoint of reader behavior that "articles on a certain topic are likely to be viewed simultaneously with articles on other topics" is easily overlooked when recommending article content based only on similarity.

[0014] On the other hand, it is also possible to make recommendations based only on the similarity of behavior regarding article contents, i.e., "articles on a certain topic are likely to be viewed simultaneously with articles on other topics." However, this method has a problem (cold start problem) in that appropriate recommendations cannot be made until a certain amount of reader behavior regarding articles is accumulated.

[0015] Therefore, the inventors of the present invention focused on a recommendation system that takes into account not only the "similarity" of the content of articles, but also the diverse behavior of readers that "likely to be viewed together." As a result, according to the proposed technology of the present invention, by using machine learning, a recommendation system is provided that takes into account not only the "similarity" of the content of articles, but also the diverse behavior of readers that "likely to be viewed together."

[0016] Thus, compared to conventional content recommendation technologies, the proposed technology of the present invention is new in that it approaches from both the article text and reader behavior, and it can be said that it optimizes the process of recommending articles to users (readers). In addition, the proposed technology of the present invention can also achieve the objective of improving the user's browsing convenience of article content. Below, one embodiment of the process related to the proposed technology of the present invention will be described in detail.

[0017] 2. System Configuration Fig. 1 is a diagram showing an example of a system according to an embodiment. Fig. 1 shows a system 1 as an example of a system according to an embodiment. Information processing according to the embodiment (i.e., processing related to the proposed technique of the present invention) is realized in the system 1.

[0018] 1, the system 1 includes a user device 10, a learning device 30, and a recommendation device 100. The user device 10, the learning device 30, and the recommendation device 100 are connected to each other via a network N so as to be able to communicate with each other via wired or wireless communication. The learning device 30 and the recommendation device 100 operate according to a program according to an embodiment (hereinafter, may be referred to as a "program PG").

[0019] The user device 10 is an information processing terminal used by a user U as a reader who browses article contents, and corresponds to an edge computer. For example, the user device 10 is a smartphone, a wearable device, a tablet terminal, a notebook PC (Personal Computer), a desktop PC, a mobile phone, a PDA (Personal Digital Assistant), or the like.

[0020] An environment (application) that allows the user device 10 to view articles and access the recommendation device 100 is installed in the user device 10. The application here may be a general-purpose application such as a browser, or a dedicated application that is newly implemented to allow access to the recommendation device 100.

[0021] The learning device 30 is a server device in charge of learning and generating a prediction model, which will be described later, and corresponds to a cloud computer. For example, the learning device 30 generates a prediction model based on learning data in which cluster information obtained by clustering article data using a predetermined clustering model is used as a target variable and feature information obtained from the article data is used as an explanatory variable, in accordance with the program PG.

[0022] The recommendation device 100 is a server device in charge of calculation processing of the belonging probability, calculation processing of the similarity, and processing of extracting article data to be recommended to a user U, and also corresponds to a cloud computer. For example, the recommendation device 100 calculates the belonging probability that the article data browsed by the user U belongs to each article cluster according to the program PG, and calculates the similarity between the belonging probability for each article cluster calculated for other article data and the belonging probability for each article cluster calculated for the article data browsed by the user. Then, the recommendation device 100 extracts article data to be recommended to the user from the other article data based on the similarity.

[0023] 1, the information processing device according to the embodiment corresponds to a series of system devices in which the learning device 30 and the recommendation device 100 work together. On the other hand, the information processing device according to the embodiment may be executed as one server device in which the learning device 30 and the recommendation device 100 are combined.

[0024] [3. Overall picture of information processing] Next, an overview of the information processing will be described with reference to Fig. 2. Fig. 2 is a diagram showing an overview of the information processing according to the embodiment. As shown in Fig. 2, the information processing according to the embodiment is divided into a learning phase and a recommendation phase. In the learning phase, the learning device 30 learns a prediction model, and in the recommendation phase, the recommendation device 100 performs a series of recommendation processes using the prediction model.

[0025] (Learning Phase) First, the learning phase will be described. In the learning phase, the learning device 30 inputs article data into a soft clustering model CM1 (an example of a first clustering model) to generate learning data in which the output information output by the soft clustering model CM1 is used as a target variable, and the article data is input into a text quantification model QM to generate learning data in which the output information output by the text quantification model QM is used as an explanatory variable. Then, the learning device 30 trains a machine learning model on this learning data to generate a prediction model PM for predicting the probability that the article data belongs to each article cluster. This will be described more specifically.

[0026] The soft clustering model CM1 is a clustering algorithm that allows article data to belong to multiple article clusters. For example, in the case of document classification, the soft clustering model CM1 is an algorithm suitable for document clustering. As an example, the soft clustering model CM1 may be PLSA (Probabilistic Latent Semantic Analysis).

[0027] 2, the learning device 30 can use article group browsing data that has been browsed by various users U as the input article group data DA to be input to the soft clustering model CM1. The article group browsing data includes, as a browsing history, article data browsed by each user U identified by a user ID.

[0028] When article group data DA (article group browsing data) is input, the soft clustering model CM1 groups article data that are likely to be read together (at the same time) by one user U or by multiple users U. Specifically, the soft clustering model CM1 creates article groups (article clusters) that group together article data that are likely to be browsed together, and calculates the probability that each piece of article data included in the article data group DA belongs to each article group. FIG. 2 shows an example in which the soft clustering model CM1 generates article groups AG1, AG2, etc. In this example, the soft clustering model CM1 calculates the probability that each piece of article data included in the article group data DA belongs to each of the article groups AG1, AG2, etc.

[0029] Next, the text quantification model QM is, for example, a distributed representation algorithm for extracting text vectors from article data. In this embodiment, Doc2Vec or the like can be used as the distributed representation algorithm. For example, as shown in FIG. 2, the learning device 30 can use article group text data as the input article group data DA to be input to the text quantification model QM. The article group text data may be composed of, for example, some of the article data included in the article group browsing data.

[0030] When article group data DA (article group sentence data) is input, the sentence quantification model QM performs a process of quantifying sentences for each article data included in the article group data DA. Specifically, the sentence quantification model QM extracts sentence vectors from each article data included in the article group data DA. Figure 2 shows an example in which the sentence quantification model QM extracts sentence vectors from each of article data AT1, article data AT2, etc.

[0031] In this state, the learning device 30 generates learning data using the probability of belonging to each article group calculated for each article data included in the article group data DA (article group browsing data) as the objective variable and the sentence vector extracted from each article data included in the article group data DA (article group sentence data) as the explanatory variable.

[0032] Then, as shown in FIG. 2, the learning device 30 inputs a combination of the objective variable (belonging probability) and explanatory variable (sentence vector) for each article data into the machine learning model as a learning data set, thereby generating a prediction model PM that predicts articles that are likely to be read together. The machine learning model to which the learning data is input learns the relationship between the explanatory variables and the objective variable. For example, when article data is input, the machine learning model learns to output the belonging probability that the input article data belongs to each article group (created by the soft clustering model CM1). As a result, the prediction model PM is obtained as a trained machine learning model. Then, the learning device 30 provides the prediction model PM to the recommendation device 100.

[0033] (Recommendation phase) Next, the recommendation phase will be described. In the recommendation phase, the recommendation device 100 uses the prediction model PM to calculate the belonging probability of the article data to be viewed by the accessed user U to each article group (created by the soft clustering model CM1). Then, the recommendation device 100 calculates the similarity between the belonging probability for each article group calculated for the article data to be viewed by the user U and the belonging probability for each article group calculated for other article data, and extracts article data to be recommended to the user U from among the other article data based on the calculated similarity. A more specific description will be given.

[0034] FIG. 2 shows an example in which a user U accesses the recommendation device 100 using a user device 10 to view article data ATx. In this example, the recommendation device 100 acquires feature information (e.g., a sentence vector) from the article data ATx to be viewed (an example of second document data), and inputs the acquired feature information to the prediction model PM. As a result, the prediction model PM outputs information regarding the probability that the article data ATx belongs to each article group. Therefore, the recommendation device 100 calculates the probability that the article data ATx belongs to each article group based on the output information by the prediction model PM.

[0035] Figure 2 shows an example in which the recommendation device 100 calculates the following probability: "80" that the article data ATx belongs to article group AG1, "35" that the article data ATx belongs to article group AG2, "20" that the article data ATx belongs to article group AG3, and "20" that the article data ATx belongs to article group AG4.

[0036] Also, it is assumed that the recommendation device 100 has already calculated the probability of each of the article data AT13, AT14, and AT15 belonging to each article group as the article data of the recommendation candidates (an example of first document data). The recommendation device 100 can also calculate the probability of each of the article data of the recommendation candidates belonging to each article group using a similar method (that is, a method using output information when feature information (e.g., a sentence vector) is input to the prediction model PM).

[0037] 2 shows an example in which the recommendation device 100 calculates the following probability: the article data AT13, which is a recommendation candidate, belongs to the article group AG1 at "80", the article data ATx belongs to the article group AG2 at "30", the article data ATx belongs to the article group AG3 at "20", and the article data ATx belongs to the article group AG4 at "20". The article data AT14 and the article data AT15 are also as shown in the figure, so their explanation will be omitted.

[0038] Next, the recommendation device 100 calculates the similarity between the belonging probability (80, 35, 20, 20) for each article group calculated for the article data ATx to be browsed and the belonging probability (80, 30, 20, 20) for each article group calculated for the article data AT13 of the recommendation candidate. For example, the recommendation device 100 calculates the similarity between the distribution state of the belonging probability (80, 35, 20, 20) among the article groups AG1 to AG4 and the distribution state of the belonging probability (80, 30, 20, 20) among the article groups AG1 to AG4.

[0039] Furthermore, the recommendation device 100 calculates the similarity between the belonging probability (80, 35, 20, 20) for each article group calculated for the article data ATx to be viewed and the belonging probability (70, 10, 0, 5) for each article group calculated for the article data AT14 of the recommendation candidate. For example, the recommendation device 100 calculates the similarity between the distribution state of the article groups AG1 to AG4, which is the belonging probability (80, 35, 20, 20), and the distribution state of the article groups AG1 to AG4, which is the belonging probability (70, 10, 0, 5).

[0040] Furthermore, the recommendation device 100 calculates the similarity between the belonging probability (80, 35, 20, 20) for each article group calculated for the article data ATx to be browsed and the belonging probability (0, 100, 10, 10) for each article group calculated for the article data AT15 of the recommendation candidate. For example, the recommendation device 100 calculates the similarity between the distribution state of the belonging probability (80, 35, 20, 20) among the article groups AG1 to AG4 and the distribution state of the belonging probability (0, 100, 10, 10) among the article groups AG1 to AG4.

[0041] 2 shows an example in which the recommendation device 100 calculates a similarity of "7.0" between the article data ATx to be viewed and article data AT13 that is a recommendation candidate. Also, an example is shown in which the recommendation device 100 calculates a similarity of "2.0" between the article data ATx to be viewed and article data AT14 that is a recommendation candidate. Also, an example is shown in which the recommendation device 100 calculates a similarity of "1.0" between the article data ATx to be viewed and article data AT15 that is a recommendation candidate.

[0042] In such a state, the recommendation device 100 may extract predetermined article data from the article data of the recommendation candidates as article data to be recommended to the access source user U. For example, the recommendation device 100 may extract a predetermined number of top-ranked article data with high similarity from the article data of the recommendation candidates as article data to be recommended to the access source user U. According to the example of FIG. 2, the recommendation device 100 can extract the article data AT13 having the highest similarity equal to or higher than a predetermined value as the article data to be recommended. In this case, the recommendation device 100 may display the article data ATx on the user device 10 and, for example, may also display link information of the article data AT13 on the user device 10 at the same time. For example, the recommendation device 100 may display a screen of the article data ATx including the link information of the article data AT13 as recommendation information on the user device 10. As another example, the recommendation device 100 may extract, from among the article data of recommendation candidates, article data having a similarity equal to or greater than a predetermined value as article data to be recommended to the user U who is the access source.

[0043] [4. Explanatory variables] 2 shows an example in which the learning device 30 generates learning data in which the sentence vectors for each article data extracted by the sentence quantification model QM are used as explanatory variables as a result of inputting article group data DA (article group sentence data) into the sentence quantification model QM. However, the learning device 30 may generate learning data in which the output information output by the soft clustering model CM2 is used as an additional explanatory variable by inputting words contained in the article data into a soft clustering model CM2 (an example of a second clustering model). FIG. 3 is an explanatory diagram for explaining explanatory variables in the learning phase in detail.

[0044] The part enclosed by the dotted line in Fig. 3 corresponds to the part that will be newly explained this time. Specifically, it corresponds to the explanation of the process of obtaining new explanatory variables by inputting words contained in article data into the soft clustering model CM2. Meanwhile, the part other than the part enclosed by the dotted line in Fig. 3 is the same as Fig. 2, so the explanation will be omitted.

[0045] For example, the learning device 30 can use article group sentence data as the input article group data DA to be input to the soft clustering model CM2 as shown in Fig. 3. The article group sentence data includes words and the frequency of occurrence of words for each piece of article data. The words referred to here are, for example, words included in sentences indicating the title of the article data, sentences indicating the main text of the article data, etc.

[0046] When article group data DA (article group sentence data) is input, the soft clustering model CM2 groups words that tend to appear together among the article data included in the article group data DA. For example, the soft clustering model CM2 creates topic groups (topic clusters) of similar word groups by grouping together words that tend to appear together among the article data included in the article group data DA. Then, the soft clustering model CM2 calculates the probability that a word belongs to each topic group. For this reason, the soft clustering model CM2 is an algorithm different from the soft clustering model CM1, and can be said to be an algorithm suitable for word clustering. As an example, the soft clustering model CM2 may be LDA (Latent Dirichlet Allocations).

[0047] 3 shows an example in which the soft clustering model CM2 generates topic groups TG1, TG2, etc. In this example, the soft clustering model CM2 calculates the probability that a word belongs to each of the topic groups TG1, TG2, etc.

[0048] 2, the learning device 30 generates learning data with sentence vectors extracted from each piece of article data included in the article group data DA (article group sentence data) as explanatory variables. Therefore, the learning device 30 calculates the belonging probability for each topic group for each piece of article data included in the article group data DA based on the word classification results by the soft clustering model CM2, combines the calculated belonging probability with the sentence vector, and generates learning data with this as an explanatory variable.

[0049] For example, the learning device 30 generates learning data with the explanatory variables being a combination of a sentence vector (extracted by the sentence quantification model QM) extracted from article data AT1 and the probability that article data AT1 belongs to each of topic groups TG1, TG2, etc. (calculated based on the classification results of words included in article data AT1).As another example, the learning device 30 generates learning data with the explanatory variables being a combination of a sentence vector (extracted by the sentence quantification model QM) extracted from article data AT2 and the probability that article data AT2 belongs to each of topic groups TG1, TG2, etc. (calculated based on the classification results of words included in article data AT2).

[0050] In this way, the learning device 30 can generate a highly accurate prediction model PM by using information that words that tend to appear together belong to the same group as explanatory variables (feature information) in the learning data.

[0051] Furthermore, the learning device 30 may analyze the article data to search for important words from among the words included in the article data, and further add information based on the search results for the important words as explanatory variables. In the above example, the learning device 30 searches for important words based on the tf-idf values ​​of the words included in the article data AT1, and further adds the tf-idf values ​​of the important words obtained by the search as explanatory variables. Furthermore, the learning device 30 searches for important words based on the tf-idf values ​​of the words included in the article data AT2, and further adds the tf-idf values ​​of the important words obtained by the search as explanatory variables.

[0052] As described above, the recommendation device 100 uses the soft clustering model CM1 to create article groups and the soft clustering model CM2 to create topic groups. On the other hand, for example, if there is one algorithm that can support both the soft clustering model CM1 and the soft clustering model CM2, the recommendation device 100 may use this algorithm to create article groups and topic groups.

[0053] 5. Configuration of the learning device The learning device 30 according to the embodiment will be described with reference to Fig. 4. Fig. 4 is a diagram showing a configuration example of the learning device 30 according to the embodiment. As shown in Fig. 4, the learning device 30 has a communication unit 31, a storage unit 32, and a control unit 33.

[0054] (Communications Department 31) The communication unit 31 is realized by, for example, a network interface card (NIC) etc. For example, the communication unit 31 is connected to the network N by wire or wirelessly, and transmits and receives information between the user device 10 and the recommendation device 100.

[0055] (Storage unit 32) The storage unit 32 is realized by, for example, a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 32 may include an article data storage unit 32a, a learning data storage unit 32b, and a prediction model data storage unit 32c.

[0056] The article data storage unit 32a stores article data to be input to the soft clustering model CM and article data to be input to the text quantification model QM. The learning data storage unit 32b stores teacher data used for learning the machine learning model. The prediction model data storage unit 32c stores information on the prediction model PM.

[0057] (Control unit 33) The control unit 33 is realized by a CPU (Central Processing Unit), MPU (Micro Processing Unit), etc., executing various programs (e.g., program PG) stored in a storage device inside the learning device 30 using a RAM as a working area. The control unit 33 is also realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0058] As shown in Fig. 4, the control unit 33 has an acquisition unit 33a, a generation unit 33b, a learning unit 33c, and a provision unit 33d, and realizes or executes the functions and actions of information processing described below. Note that the internal configuration of the control unit 33 is not limited to the configuration shown in Fig. 4, and may be other configurations as long as they perform information processing described below. Also, the connection relationship between the processing units of the control unit 33 is not limited to the connection relationship shown in Fig. 4, and may be other connection relationships.

[0059] (Acquisition part 33a) The acquisition unit 33a acquires information necessary for generating learning data and machine learning. For example, the acquisition unit 33a acquires article data to be input to the soft clustering model CM and article data to be input to the text quantification model QM. The acquisition unit 33a also acquires learning data to be input to the machine learning model.

[0060] (Generation unit 33b) The generating unit 33b generates learning data to be input to the machine learning model. For example, the generating unit 33b generates learning data in which cluster information obtained by clustering input article data by the soft clustering model CM1 is used as a target variable, and feature information obtained from the input document data is used as an explanatory variable.

[0061] For example, the generation unit 33b generates learning data using the probability of belonging to each article cluster (a cluster according to the viewing tendency of the article data) calculated for the input article data as the objective variable and vector information extracted from the input article data as the explanatory variable.

[0062] The generating unit 33b may also search for important words from among the words included in the input document data by analyzing the input article data. The generating unit 33b may then generate learning data to which information based on the search results for important words (e.g., tf-idf scores) is added as explanatory variables.

[0063] Furthermore, the generating unit 33b may generate learning data to which cluster information obtained by clustering words included in the input article data by the soft clustering model CM2 is added as an explanatory variable. For example, the generating unit 33b generates learning data to which the belonging probability of each topic cluster (cluster according to the appearance tendency of the word) calculated for the words included in the input article data is added as an explanatory variable.

[0064] (Learning section 33c) The learning unit 33c generates a prediction model by machine learning based on the learning data generated by the generation unit 33b. For example, the learning unit 33c generates a prediction model that predicts articles that are likely to be read together by inputting a combination of a target variable (belonging probability) for each article data and an explanatory variable (feature information such as a sentence vector) for each article data into the machine learning model as a learning data set.

[0065] The machine learning model to which the learning data is input learns the relationship between explanatory variables and objective variables. For example, when article data is input, the machine learning model learns to output the probability that the input article data belongs to each article group (created by the soft clustering model CM1).

[0066] (Provider 33d) The providing unit 33d provides the trained prediction model to the recommendation device 100.

[0067] [6. Example of generating learning data] Next, a description will be given of generation of learning data performed by the learning device 30. Fig. 5 is a diagram showing an example of generation of learning data according to an embodiment. In Fig. 5, the contents described in Fig. 2 will be described in more detail.

[0068] 5 also shows a scene in which learning data is generated from one piece of article data AT1 among the article data included in the article group data DA. Note that, although attention is focused on the article data AT1 in this way in FIG. 5, learning data is generated in the same manner for other article data (e.g., article data AT2, etc.) included in the article group data DA.

[0069] First, the objective variable will be described. The generation unit 33b inputs the article group data DA including the article data AT1 to the soft clustering model CM1. Specifically, the generation unit 33b inputs the article group data DA included in the browsing history of each user U to the soft clustering model CM1.

[0070] When the article group data DA is input, the soft clustering model CM1 creates an article group that groups together article data that are likely to be viewed together, and calculates the probability that each piece of article data included in the article group data DA belongs to each article group. Fig. 5 shows an example in which the soft clustering model CM1 generates article groups AG1, AG2, AG3, and AG4. Fig. 5 also shows an example in which the soft clustering model CM1 calculates the following probability of the article data AT1 belonging to the article group AG1: "80", the probability of the article data AT1 belonging to the article group AG2: "35", the probability of the article data AT1 belonging to the article group AG3: "20", and the probability of the article data AT1 belonging to the article group AG4: "20".

[0071] In this example, the generating unit 33b determines the belonging probability (80, 35, 20, 20) for each article group calculated for the article data AT1 as the objective variable TV in the learning data corresponding to the article data AT1.

[0072] Next, the explanatory variables will be described. The generation unit 33b inputs the article group data DA including the article data AT1 to the text quantification model QM. When the article group data DA is input, the text quantification model QM acquires feature information from each article data included in the article group data DA. Figure 5 shows an example in which the text quantification model QM extracts a text vector as feature information from the article data AT1.

[0073] Moreover, the generating unit 33b inputs the words constituting the article group data DA including the article data AT1 to the soft clustering model CM2.

[0074] The soft clustering model CM2 creates topic groups of similar words by grouping together words that tend to appear together in article data included in the article group data DA based on the frequency of occurrence of the words included in the article group data DA. The soft clustering model CM2 then calculates the probability that a word belongs to each topic group. Figure 5 shows an example in which the soft clustering model CM2 generates topic groups TG1, TG2, TG3, and TG4.

[0075] In this state, the generation unit 33b calculates the probability of the article data AT1 belonging to each of the topic groups TG1 to TG4 based on the classification result obtained by classifying each word included in the article data AT1 by the soft clustering model CM2. For example, the generation unit 33b can calculate the probability of the article data AT1 belonging to each of the topic groups TG1 to TG4 by scoring using the probability of belonging for each topic group calculated for words included in the article data AT1 whose occurrence frequency is equal to or higher than a predetermined value. FIG. 5 shows an example in which the soft clustering model CM2 calculates the probability of the article data AT1 belonging to the topic group TG1 as "80", the probability of the article data AT1 belonging to the topic group TG2 as "20", the probability of the article data AT1 belonging to the topic group TG3 as "5", and the probability of the article data AT1 belonging to the topic group TG4 as "5".

[0076] Here, the generation unit 33b defines the combination of the sentence vector extracted as feature information from the article data AT1 and the belonging probability (80, 20, 5, 5) for each topic group calculated for the article data AT1 as the explanatory variable EV in the learning data corresponding to the article data AT1.

[0077] Then, the generating unit 33b inputs the learning data configured of a pair of the objective variable TV and the explanatory variable EV to the prediction model PM, as shown in FIG.

[0078] 5, when the prediction model PM is trained using the article affiliation probability for an article group as the objective variable and feature information, which is a pair of the article affiliation probability for a topic group and a sentence vector, as the explanatory variable, the recommendation device 100 operates as follows: Specifically, the recommendation device 100 inputs a pair of the sentence vector acquired from the article data ATx to be viewed and the affiliation probability of the article data ATx for the topic group into the prediction model PM, and calculates the affiliation probability that the article data ATx belongs to each article group from the output result.

[0079] 7. Configuration of the recommendation device The recommendation device 100 according to the embodiment will be described with reference to Fig. 6. Fig. 6 is a diagram showing an example of the configuration of the recommendation device 100 according to the embodiment. As shown in Fig. 6, the recommendation device 100 includes a communication unit 110, a storage unit 120, and a control unit 130.

[0080] (Communication unit 110) The communication unit 110 is realized by, for example, a NIC etc. For example, the communication unit 110 is connected to the network N by wire or wirelessly, and transmits and receives information between the user device 10 and the learning device 30.

[0081] (Storage unit 120) The storage unit 120 is realized by, for example, a semiconductor memory element such as a RAM or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 120 may include a prediction model data storage unit 121, an association probability data storage unit 122, and a similarity data storage unit 123.

[0082] The prediction model data storage unit 121 stores information on the prediction model PM. The belonging probability data storage unit 122 stores information on the belonging probability. The similarity data storage unit 123 stores data on the similarity.

[0083] (Control unit 130) The control unit 130 is realized by a CPU, an MPU, or the like, executing various programs (for example, a program PG) stored in a storage device inside the recommendation device 100 using a RAM as a working area. The control unit 130 is also realized by an integrated circuit such as an ASIC or an FPGA.

[0084] As shown in Fig. 6, the control unit 130 has a reception unit 131, a probability calculation unit 132, an acquisition unit 133, a similarity calculation unit 134, an extraction unit 135, and a recommendation unit 136, and realizes or executes the functions and actions of information processing described below. Note that the internal configuration of the control unit 130 is not limited to the configuration shown in Fig. 6, and may be other configurations as long as they perform the information processing described below. Also, the connection relationship between the processing units of the control unit 130 is not limited to the connection relationship shown in Fig. 6, and may be other connection relationships.

[0085] (Reception Department 131) The receiving unit 131 receives a request to view article content from the user U. For example, the receiving unit 131 receives from the user device 10 a request to view article content corresponding to information (for example, a URL) specifying article data to be viewed.

[0086] (Probability calculation unit 132) The probability calculation unit 132 calculates the probability that the article data belongs to each article group using the prediction model PM generated based on the clustering model CM1. For example, the probability calculation unit 132 calculates the probability that the article data of the recommendation candidate (an example of first document data) belongs to each article group. In addition, the probability calculation unit 132 calculates the probability that the article data to be browsed by the user U (an example of second document data) belongs to each article group.

[0087] For example, the probability calculation unit 132 inputs feature information obtained from the article data to the prediction model PM, and calculates the probability that the document data belongs to each article group based on the output information output from the prediction model PM. The feature information here may be a sentence vector extracted from the article data, the article data's belonging probability to each topic group, a tf-idf value calculated for important words included in the article data, etc.

[0088] (Acquisition part 133) The acquisition unit 133 acquires information necessary for recommendation. For example, the acquisition unit 133 acquires information necessary for similarity calculation. For example, the acquisition unit 133 acquires the belonging probability for each article group calculated in advance for the article data of the recommendation candidate corresponding to the article data to be browsed by the user U. The belonging probability for each article group calculated in advance for the article data of the recommendation candidate may be stored in the storage unit 120.

[0089] (Similarity calculation unit 134) The similarity calculation unit 134 calculates the similarity between the belonging probability for each article group calculated for the article data of the recommendation candidate and the belonging probability for each article group calculated for the article data of the browsing target. For example, the similarity calculation unit 134 may calculate the similarity between the distribution state between article groups of the belonging probability calculated for the article data of the recommendation candidate and the distribution state between article groups of the belonging probability calculated for the article data of the browsing target based on a cosine similarity calculation algorithm.

[0090] (Extraction part 135) The extraction unit 135 extracts article data to be recommended to the user U from the article data of the recommendation candidates based on the similarity calculated by the similarity calculation unit 134. For example, the extraction unit 135 may extract a top predetermined number of article data with high similarity from the article data of the recommendation candidates as article data to be recommended to the user U. As another example, the extraction unit 135 may extract article data with a similarity equal to or greater than a predetermined value from the article data of the recommendation candidates as article data to be recommended to the user U. The number of article data to be recommended is not limited.

[0091] (Recommendation Section 136) The recommendation unit 136 provides the recommended article data extracted by the extraction unit 135 to the user U. For example, the recommendation unit 136 may cause the user device 10 to display the article data to be browsed, and may also cause the user device 10 to simultaneously display link information of the recommended article data.

[0092] 8. Processing Procedure Next, the procedure of information processing according to the embodiment will be described with reference to Fig. 7 and Fig. 8. Fig. 7 describes the processing procedure of the learning device 30 that operates according to the program PG, and Fig. 8 describes the processing procedure of the recommendation device 100 that operates according to the program PG.

[0093] 8-1. Processing procedure of the learning device 7 is a flowchart explaining the operation procedure of the learning device 30. The generation unit 33b generates learning data (step S701). The learning data is generated using cluster information obtained by clustering input article data as a target variable and feature information obtained from the input document data as an explanatory variable.

[0094] The learning unit 33c generates a prediction model based on the learning data (step S702). For example, the learning unit 33c generates the prediction model by training a model so as to output the probability that the input article data belongs to each article group (created by the soft clustering model CM1) when the article data is input.

[0095] The providing unit 33d provides the prediction model to the recommendation server 100 (Step S703).

[0096] 8-2. Processing procedure of the recommendation device Fig. 8 is a flowchart illustrating the operation procedure of the recommendation device 100. In the example of Fig. 8, it is assumed that the probability calculation unit 132 has calculated in advance the belonging probability for each article group for the article data of various recommendation candidates.

[0097] The reception unit 131 determines whether or not a viewing request for requesting article data to be viewed has been received from the user device 10 of the user U (step S801). If the reception unit 131 has not received a viewing request (step S801; No), the reception unit 131 waits until a viewing request is received.

[0098] When a viewing request is accepted (step S801; Yes), the probability calculation unit 132 uses a prediction model to calculate the belonging probability that the article data to be viewed belongs to each article group (step S802).

[0099] The acquiring unit 133 also acquires the calculated belonging probability for each article group for the article data of the recommendation candidate corresponding to the article data to be browsed (step S803).

[0100] Then, the similarity calculation unit 134 calculates the similarity between the belonging probability for each article group calculated for the article data of the recommendation candidate and the belonging probability for each article group calculated for the article data of the browsing target (step S804).

[0101] The extraction unit 135 extracts article data to be recommended to the user U who has made the browsing request from among the article data of recommendation candidates corresponding to the article data to be browsed based on the similarity (step S805). The recommendation unit 136 executes a recommendation process to provide the extracted article data to the user U (step S806).

[0102] [9. Hardware Configuration] Next, a hardware configuration example of an information processing device (learning device 30, recommendation device 100) according to an embodiment will be described. FIG. 9 is a block diagram showing a hardware configuration example of an information processing device according to an embodiment. Referring to FIG. 9, the information processing device according to an embodiment includes, for example, a processor 801, a ROM 802, a RAM 803, a host bus 804, a bridge 805, an external bus 806, an interface 807, an input device 808, an output device 809, a storage 810, a drive 811, a connection port 812, and a communication device 813. Note that the hardware configuration shown here is an example, and some of the components may be omitted. In addition, the information processing device may further include components other than those shown here.

[0103] (Processor 801) The processor 801 functions, for example, as an arithmetic processing device or control device, and controls all or part of the operation of each component based on various programs recorded in the ROM 802, the RAM 803, the storage 810, or the removable recording medium 901.

[0104] (ROM802, RAM803) The ROM 802 is a means for storing the programs to be read into the processor 801, data to be used for calculations, etc. The RAM 803 temporarily or permanently stores, for example, the programs to be read into the processor 801, various parameters that change appropriately when the programs are executed, etc.

[0105] (host bus 804, bridge 805, external bus 806, interface 807) The processor 801, ROM 802, and RAM 803 are connected to each other via, for example, a host bus 804 capable of high-speed data transmission. On the other hand, the host bus 804 is connected to an external bus 806 having a relatively low data transmission speed via, for example, a bridge 805. In addition, the external bus 806 is connected to various components via an interface 807.

[0106] (Input device 808) Examples of the input device 808 include a mouse, a keyboard, a touch panel, a button, a switch, and a lever. Furthermore, a remote controller capable of transmitting control signals using infrared rays or other radio waves may be used as the input device 808. The input device 808 also includes an audio input device such as a microphone.

[0107] (Output Device 809) The output device 809 is a device capable of visually or audibly notifying the user of acquired information, such as a display device such as a CRT (Cathode Ray Tube), LCD, or organic EL, an audio output device such as a speaker or a headphone, a printer, a mobile phone, or a facsimile. The output device 809 according to this embodiment also includes various vibration devices capable of outputting tactile stimulation. The output device 809 may also include an AI speaker or a wearable device.

[0108] (Storage 810) The storage 810 is a device for storing various types of data. For example, a magnetic storage device such as a hard disk drive (HDD), a semiconductor storage device, an optical storage device, or a magneto-optical storage device may be used as the storage 810.

[0109] (Drive 811) The drive 811 is a device that reads information recorded on a removable recording medium 901 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and writes information to the removable recording medium 901.

[0110] (Connection port 812) The connection port 812 is a port for connecting an external device 902, such as a Universal Serial Bus (USB) port, an IEEE1394 port, a Small Computer System Interface (SCSI), an RS-232C port, or an optical audio terminal.

[0111] (Communication device 813) The communication device 813 is a communication device for connecting to a network, such as a communication card for wired or wireless LAN, Bluetooth (registered trademark), or WUSB (Wireless USB), a router for optical communications, a router for ADSL (Asymmetric Digital Subscriber Line), or a modem for various types of communications.

[0112] (Removable recording media 901) The removable recording medium 901 is, for example, a DVD medium, a Blu-ray (registered trademark) medium, an HD DVD medium, various semiconductor storage media, etc. Of course, the removable recording medium 901 may be, for example, an IC card equipped with a non-contact type IC chip, or an electronic device.

[0113] (External connection device 902) The external connection device 902 is, for example, a printer, a portable music player, a digital camera, a digital video camera, or an IC recorder.

[0114] [10. Other] Among the above-mentioned processes, all or part of the processes described as being performed automatically may be performed manually. Also, all or part of the processes described as being performed manually may be performed automatically by a known method. In addition, the information including the processing procedures, specific names, various data and parameters shown in the above documents and drawings may be changed arbitrarily unless otherwise specified. For example, the various information shown in each drawing is not limited to the illustrated information.

[0115] In addition, each component of each device shown in the figure is a functional concept, and does not necessarily have to be physically configured as shown in the figure. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure. Furthermore, each component may be configured by distributing and integrating functionally or physically in any unit, in whole or in part, depending on various loads and usage conditions. Furthermore, each process described above may be executed in appropriate combination within a range that does not contradict.

[0116] The above describes the embodiments of the present application in detail with reference to several drawings. However, these are merely examples, and the present invention can be embodied in other forms that incorporate various modifications and improvements based on the knowledge of those skilled in the art, including the aspects described in the Disclosure of the Invention section. [Explanation of symbols]

[0117] 1 System 10 User Device 30 Learning Device 33a Acquisition part 33b Generator 33c Learning Department 33d supply department 100 Recommendation Device 131 Reception 132 Probability Calculation Unit 133 Acquisition Department 134 Similarity calculation unit 135 Extraction part 136 Recommendation Department

Claims

1. an acquisition step of acquiring a probability that the first document data belongs to each of the document clusters created by inputting the document data to a first clustering model, the document clusters being determined based on the browsing tendency of the document data; a probability calculation step of calculating a probability that second document data browsed by a user belongs to each of the document clusters by using a prediction model generated based on the first clustering model; a similarity calculation step of calculating a similarity between the belonging probability for each of the document clusters calculated for the first document data and the belonging probability for each of the document clusters calculated for the second document data; an extraction step of extracting document data to be recommended to the user from the first document data based on the similarity; A program for causing a computer to execute the following.

2. further executing a generation step of generating the prediction model based on learning data having cluster information obtained by clustering input document data by the first clustering model as a target variable and feature information obtained from the input document data as an explanatory variable; The probability calculation step includes inputting feature information obtained from the second document data to the prediction model, and calculating a probability that the second document data belongs to each of the document clusters based on output information output from the prediction model.

2. The program according to claim 1 .

3. the cluster information includes a probability that the input document data belongs to each of the document clusters; The generation step generates the prediction model based on the learning data having the probability of belonging to each document cluster calculated for the input document data as a response variable and vector information extracted from the input document data as an explanatory variable.

3. The program according to claim 2 .

4. further executing a search procedure for searching for important words from among words included in the input document data by analyzing the input document data; The generation step generates the prediction model based on the learning data having information based on the search results of the important words as explanatory variables.

3. The program according to claim 2 .

5. The acquisition step includes acquiring a probability that the input document data belongs to each topic cluster created by inputting words contained in the input document data to a second clustering model according to an appearance tendency of the words, and The generation step generates the prediction model based on the learning data in which the probability of belonging to each word cluster calculated for the input document data is used as an explanatory variable.

3. The program according to claim 2 .

6. The similarity calculation step calculates a similarity between a distribution state of the affiliation probability calculated for the first document data among the document clusters and a distribution state of the affiliation probability calculated for the second document data among the document clusters.

2. The program according to claim 1 .

7. The extraction unit extracts a top predetermined number of document data having high similarity from the first document data as document data to be recommended to the user.

2. The program according to claim 1 .

8. The extraction step includes extracting, from the first document data, document data having a similarity equal to or greater than a predetermined value as document data to be recommended to the user.

2. The program according to claim 1 .

9. an acquisition unit that acquires a probability that the first document data belongs to each of document clusters created by inputting document data to a first clustering model, the document clusters being in accordance with a browsing tendency of the document data; a probability calculation unit that calculates a probability that second document data browsed by a user belongs to each of the document clusters by using a prediction model generated based on the first clustering model; a similarity calculation unit that calculates a similarity between the belonging probability for each of the document clusters calculated for the first document data and the belonging probability for each of the document clusters calculated for the second document data; an extraction unit that extracts document data to be recommended to the user from the first document data based on the similarity; 13. An information processing device comprising:

10. An information processing method executed by an information processing device, an acquisition step of acquiring a probability that the first document data belongs to each of the document clusters created by inputting the document data to a first clustering model, the document clusters being determined based on the browsing tendency of the document data; a probability calculation step of calculating a probability that second document data browsed by a user belongs to each of the document clusters by using a prediction model generated based on the first clustering model; a similarity calculation step of calculating a similarity between the belonging probability for each of the document clusters calculated for the first document data and the belonging probability for each of the document clusters calculated for the second document data; an extraction step of extracting document data to be recommended to the user from the first document data based on the similarity; 13. An information processing method comprising:

Images