JP2014110004A - Information processor and information processing method and data structure - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processor capable of easily updating a directed graph constituting a taste model.SOLUTION: An information processor 10 uses a taste model 151 having a data structure in which a determination node in a directed graph includes one attribute information among a plurality of attribute information included in program information and information indicating a determination formula for determining the degree of taste of a user related to the determination node, and a start node is the start point of a program for generating one or more rules for performing predetermined determination related to the taste of the user on the basis of a determination result based on the determination formula associated with the determination node following the start node, with which connection information to one or more determination nodes is associated. The information processor 10 includes a content evaluation part 120 for estimating the taste of the user by inputting the program information, and applying one or more rules generated on the basis of the taste model 151 to the program information.

Description

  The present invention relates to an information processing apparatus, an information processing method, and a data structure that are suitable for use when recommending content that meets a user's preference.

  There is a technique for automatically extracting and recommending content that suits the user's preference using meta-information including characters, codes, and the like representing the content when viewing or recording content such as a moving image. For example, Patent Literature 1 discloses a technology that enables the selection of content that suits the user's preference automatically by using data representing the degree of preference of another user via a network.

  Patent Document 2 discloses a technique for recommending content using a preference model generated using a Bayesian Network. The preference model described in Patent Document 2 is represented by a plurality of nodes and connection information indicating a parent-child relationship between the plurality of nodes. Here, the plurality of nodes represent random variables corresponding to items included in the program information. Further, the connection between the parent node and the child node represents a causal relationship, that is, a relationship between the cause and the result. The relationship between each of these nodes is represented by a directed graph. Furthermore, the causal relationship between the parent node and the child node is expressed by a conditional probability. In other words, the child node random variable is obtained on the condition that the result obtained from the parent node random variable is used as a condition. The value of the conditional probability represented by each node corresponds to the viewing probability for each item of the user. In the system of Patent Document 2, the preference model learning unit calculates the value of the conditional probability in the preference model based on the viewing history for a certain period in the past.

JP 2003-114903 A JP 2008-262398 A

  In the technique described in Patent Document 1, the amount of information that can be used when estimating the user's preference is increased by using data indicating the degree of preference of other users. However, since it is premised on a network connection, there are certain problems regarding an increase in traffic volume and use in an offline environment.

  On the other hand, in the technique described in Patent Document 2, there is a problem that the process of adding a new item to the preference model has the following restrictions, for example. That is, the preference model described in Patent Document 2 has a parent-child relationship between a plurality of nodes. Therefore, for example, when a new node is added or deleted, the contents of the child node whose parent node is the added or deleted node must be changed. Therefore, there is a problem that the process of adding or deleting nodes becomes excessive.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide an information processing apparatus, an information processing method, and a data structure that can solve the above-described problems.

  In order to solve the above-described problem, an information processing apparatus according to the present invention is represented by a directed graph including a plurality of nodes and a plurality of edges that connect the plurality of nodes with directionality, and the plurality of nodes The node is a start node or a determination node, and the determination node includes one attribute information among the plurality of attribute information included in the content information representing the attribute of the content using the plurality of attribute information, and the determination Information indicating a determination formula for determining a user's preference degree related to a node is associated with the start node, and the start node is associated with one or more determination nodes subsequent to the start node. At the starting point of a program that generates one or a plurality of rules for making a predetermined determination regarding the user's preference based on one or a plurality of determination results based on one or a plurality of the determination formulas A storage unit storing a preference model having a data structure associated with connection information to one or a plurality of the determination nodes, and the content information is input and generated based on the preference model And an estimation unit that estimates the user's preference by applying one or more of the rules to the content information.

  Further, in another information processing apparatus of the present invention, the determination node is further associated with a predetermined parameter for representing the reliability of the determination expression associated with the determination node. In the rule generated based on the preference model, the one or more representing the reliability of the one or more determination formulas obtained from the one or more determination results used as a reference when generating the rule A value obtained by using a predetermined parameter, and a value representing the reliability of the rule is set, and when the estimation unit applies the one or more rules to the content information, A value representing each reliability set in each rule is used.

  In another information processing apparatus of the present invention, the preference model is generated or updated based on a plurality of pieces of the content information, and the predetermined parameter associated with the determination node is It is characterized in that it corresponds to the frequency of appearance of the attribute information associated with the determination node in the plurality of pieces of content information used as a reference when generating or updating a preference model.

  Further, another information processing apparatus according to the present invention is characterized in that the estimation unit estimates the user's preference while performing an evolution process on the preference model using a genetic algorithm.

  The information processing method of the present invention is represented by a directed graph including a plurality of nodes and a plurality of edges that connect the plurality of nodes with directionality, and the plurality of nodes are start nodes or A determination node, wherein the determination node includes one attribute information of the plurality of attribute information included in the content information representing the attribute of the content using the plurality of attribute information, and the user of the determination node Information indicating a determination formula for determining the degree of preference, and the start node is associated with one or a plurality of the determination nodes subsequent to the start node. A starting point of a program for generating one or more rules for making a predetermined determination related to a user's preference based on one or more determination results based on a determination formula. A preference model having a data structure associated with connection information to the determination node is stored in a predetermined storage unit, the content information is input by the estimation unit, and generated based on the preference model The user's preference is estimated by applying one or a plurality of the rules to the content information.

  The data structure of the present invention is a data structure for representing a user's preference model, and includes a plurality of nodes and a plurality of edges that connect the plurality of nodes with directionality. Represented by a directed graph, wherein the plurality of nodes are start nodes or determination nodes, and the determination node includes a plurality of pieces of attribute information included in the content information representing content attributes using the plurality of attribute information. One attribute information is associated with information representing a determination formula for determining the degree of preference of the user related to the determination node, and the start node is one or more of the one or more of the above-described start nodes One or more routes for making a predetermined determination regarding the user's preference based on one or more determination results based on one or more of the determination formulas associated with the determination node A starting point of the program that generates the connection information to one or more of the determining node is associated, characterized in that.

  In addition, another data structure of the present invention is further characterized in that a predetermined parameter for representing the reliability of the determination formula associated with the determination node is associated.

  According to the present invention, there is no need for a parent-child relationship between information associated with each determination node and information associated with other determination nodes. Therefore, it is possible to easily add or delete determination nodes from the preference model. In addition, since the information for generating or updating the preference model of the present invention can be obtained from content information such as program information, the traffic volume is increased as described above, or used in an offline environment. There is no problem of the above restrictions.

It is the block diagram which showed the structure of one Embodiment of this invention. It is a schematic diagram for demonstrating the structural example of the preference model 151 shown in FIG. It is explanatory drawing for demonstrating an example of the data showing the directed graph 151a shown in FIG. It is a schematic diagram for demonstrating the structural example of the determination node content information 151b shown in FIG. FIG. 3 is an explanatory diagram for describing specific examples of a start node 121a1 and a determination node 121a2 illustrated in FIG. 2; FIG. 6 is an explanatory diagram for describing a specific example of determination node content information 151b associated with each determination node 121a2 illustrated in FIG. 5; It is explanatory drawing for demonstrating the rule which may be produced | generated from the directed graph 151a shown in FIG. It is explanatory drawing for demonstrating the example of the rule produced | generated from the preference model 151 shown in FIG. 3 is a flowchart for explaining an operation example of the information processing apparatus 10 illustrated in FIG. 1. It is a flowchart for demonstrating the content of the content evaluation process performed by FIG.9 S103. It is explanatory drawing for demonstrating the content of the wrinkle process performed by step S202 of FIG. It is explanatory drawing for demonstrating an example of the crossover process which is one of the evolution processes performed by step S204 of FIG. It is explanatory drawing for demonstrating an example of the mutation process which is one of the evolution processes performed by step S204 of FIG. It is explanatory drawing for demonstrating the content of the preference estimation process performed by FIG.10 S205.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of an information processing system (information processing apparatus) 10 as an embodiment of the present invention. The information processing system 10 includes a content display device 100 and a content control device 200. Each configuration in the information processing system 10 illustrated in FIG. 1 is realized by a combination of one or a plurality of CPUs (central processing units), peripheral devices thereof, and programs executed by the CPUs. Can do. The content display device 100 includes a content reception unit 110, a content output unit 115, a content evaluation unit 120, a recommendation result display selection unit 130, a content browsing log generation unit 140, a preference model generation unit 150, and a training information setting unit 160. Yes.

  The content display device 100 is, for example, an information terminal such as a TV (television), a smartphone, a mobile phone, or a PC (personal computer). Alternatively, the content display device 100 may be a combination of an information terminal such as a TV and a content reception / recording device such as a set top box or a video recording device.

  The content display device 100 is a device capable of receiving and displaying content by the content receiving unit 110 and the content output unit 115. The content receiving unit 110 selects content to be received in accordance with an instruction from the recommendation result display selection unit 130, and receives the content via a predetermined medium such as a broadcast wave or a network. Then, the content receiving unit 110 outputs a signal representing the received content to the content output unit 115 by converting it into a video signal or audio signal in a predetermined format. The content output unit 115 includes a display and a speaker, and performs video display and audio output based on the input video signal and audio signal. In addition, the content receiving unit 110 receives content information (hereinafter referred to as program information) representing content attributes using a plurality of attribute information such as electronic program information, and outputs the content information to the content evaluation unit 120.

  The content evaluation unit 120 (estimation unit) evaluates the user's preference (that is, the user's preference) with respect to the received content, and creates program information list information recommended to the user. On the other hand, the preference model generation unit 150 generates / updates a preference model 151 that is a model representing a preference (likes and dislikes) regarding a user's content. The content evaluation unit 120 evaluates the content using the preference model 151 generated by the preference model generation unit 150, and uses one or more program information estimated to be suitable for the user's preference as a recommendation result of the content. The data is output to the display selection unit 130.

  The recommendation result display selection unit 130 displays the content recommendation result based on the evaluation result of the content evaluation unit 120. Further, the recommendation result display selection unit 130 displays the program information for training of the preference model 151 in accordance with the instruction of the training information setting unit 160. At this time, the recommendation result display selection unit 130 outputs, via the content receiving unit 110, the content or image representing the content recommendation result received from the content evaluation unit 120 or the program information received from the training information setting unit 160. Output from the unit 115. In addition, the recommendation result display selection unit 130 determines content to be selected based on a predetermined control signal received from the content control device 200, and outputs the determined result to the content reception unit 110. Here, as described above, the content receiving unit 110 receives the selected content via a predetermined medium, converts it into a predetermined video signal or audio signal, and outputs it from the content output unit 115.

  The content browsing log generation unit 140 generates a log relating to user preferences based on control details such as selection by the content control unit 210 for the content recommendation result displayed by the recommendation result display selection unit 130. That is, the content browsing log generation unit 140 records a plurality of pieces of program information of content that the user has instructed to display or record. The program information recorded by the content browsing log generation unit 140 is used, for example, as training program information representing content that meets the user's preference when the preference model generation unit 150 updates the preference model 151.

  The training information setting unit 160 displays training program information from the content output unit 115 via the recommendation result display selection unit 130 and the like. Further, the training information setting unit 160, based on a predetermined control signal received from the content control device 200 in response to a user operation for displaying the training program information, the user's preference (“like”, “I dislike” is set, and the set result is output to the preference model generation unit 150 together with the program information.

  The preference model generation unit 150 uses pre-prepared training program information (that is, program information in which information indicating correct evaluation results indicating likes and dislikes is associated in advance) or a content browsing log generation unit The user's preference model 151 is generated or updated using the log generated by 140 or the program information for training set by the training information setting unit 160. The preference model 151 is stored in a predetermined storage unit in the preference model generation unit 150. In addition, the preference model generation unit 150 stores training program information used when the preference model 151 is generated or updated, together with the evaluation results of likes and dislikes, in a predetermined storage unit.

  On the other hand, the content control device 200 is configured as a TV remote control, a mobile phone touch panel, or a PC mouse. The content control device 200 includes a content control unit 210 having an operation unit and a signal transmission unit. In the content control device 200, the content control unit 210 inputs a user's predetermined operation and outputs a predetermined control signal indicating the input result to the recommendation result display unit 130 and the training information setting unit 160. Specifically, the content control unit 210 performs operation information such as selection and display instructions for the user's content, selection of the user such as “like” and “dislike” for the program information for training, and setting instructions. Input and output control signals according to the contents of those operations. Note that the content control device 200 may be configured as an apparatus different from the content display device 100 as illustrated in FIG. 1, or may be configured integrally with the content display device 100.

  Here, with reference to FIGS. 2 to 8, a configuration example of the preference model 151 shown in FIG. 1 will be described. FIG. 2 is a schematic diagram showing a configuration example of the preference model 151 shown in FIG. The preference model 151 is a type that typically represents a user's preference (like, likes and dislikes) for content. Based on the preference model 151, a plurality of rules for performing a predetermined determination process when a user's preference for unknown program information is estimated is generated. The preference model 151 is a model whose contents can be updated by learning, but an initial model can be generated as follows, for example. In other words, a preference such as “like” or “dislike” is set in advance for a plurality of program information prepared in advance, or a new preference such as “like” or “dislike” is set by the user. What is requested can be generated by using it as program information for training. Here, a plurality of training program information for which preference settings are prepared in advance may be prepared in accordance with, for example, age group and sex. In addition, the preference model 151 itself is not generated from zero, but a state in which a constant training is performed can be prepared as an initial state. The program information used when creating the initial model is stored in a predetermined storage unit in the preference model generation unit 150.

  As illustrated in FIG. 2, the preference model 151 includes information indicating a directed graph 151 a indicating a connection relationship between the plurality of nodes 151 a 1 or 151 a 2 in the preference model 151 and information associated with the plurality of determination nodes 151 a 2. It is expressed by a plurality of pieces of determination node content information 151b which is data indicating the content for each node. That is, the preference model 151 includes a directed graph 151a including a plurality of nodes 151a1 or 151a2 and a plurality of edges 151a3 that connect the plurality of nodes 151a1 or 151a2 with directionality, and a plurality of pieces of determination node content information. 151b. The plurality of nodes 151a1 or 151a2 are classified into a start node 151a1 and a determination node 151a2. Based on the structure of the directed graph 151a and the information associated with the determination node 151a2, a plurality of rules are generated to make a predetermined determination regarding the user's preference.

  The start node 151a1 indicates a start point of a program (or calculation procedure) for creating a rule, and is a node connected to the determination node 151a2. The determination node 151a2 is a node associated with one piece of “attribute information” for representing a program. The “determination formula” of this node alone and the “predetermined parameter corresponding to the reliability” of the determination formula are Are associated. At this time, the attribute information for representing the program refers to information related to the program, such as information representing the characteristics of the program, information representing the nature, such as “drama”, “broadcast time”, and “channel”. In this case, the determination expression associated with the determination node 151a2 is a mathematical expression (or a program that outputs a value indicating the degree of “like and dislike” when the program information includes attribute information associated with the determination node. ), Which is a mathematical expression or condition for determining (or digitizing) the degree of preference. Each determination node can calculate the degree of “like” or “dislike” for the node alone by using this determination formula. The predetermined parameter corresponding to the reliability is a parameter for expressing the reliability of the determination formula. For example, how much attribute information associated with the determination node appears in the training program information. Calculated based on frequency. By using this “predetermined parameter corresponding to the reliability”, for example, it is assumed that the reliability is high when the frequency is high and the reliability is low when the frequency is low, and relative to the result obtained from the determination formula. Weighting can be performed. The preference model 151 uses these two types of nodes and the directed graph to construct a gene structure that generates a logical expression that represents a rule for determining liking or dislike of a program.

  In FIG. 2, the start node 151a1 is indicated by a hexagonal symbol, and the numbers “1” to “3” in the symbol represent IDs (identification codes) set to the start nodes 151a1. ing. Further, the decision node 151a2 is indicated by a round symbol, and the numbers “1” to “8” in the symbol represent the ID set for each decision node 151a2.

  The connection relationship between each start node 151a1, each determination node 151a2, and each directed edge 151a3 in the directed graph 151a illustrated in FIG. 2 can be represented by, for example, two types of tables illustrated in FIG. That is, as shown in FIG. 3A, the connection relationship between each start node 151a1 and one or more determination nodes 151a2 directly connected to each start node 151a1 is as follows. And the ID of one or more connection destination determination nodes directly connected to the start node can be represented by a table. In the table of FIG. 3A, for example, a start node 151a1 (hereinafter also referred to as “start node 1”) having a start node ID “1” is a determination node 151a2 (hereinafter “determination node”) having an ID “4”. 4 ") indicates that the connection is made directly in the direction from the start node 1 to the determination node 4. Further, the start node 151a1 having the start node ID “2” (hereinafter also referred to as “start node 2” (hereinafter the same)) has three determinations with IDs “2”, “5”, and “7”. This indicates that the node 151a2 is directly connected in the direction from the start node to each determination node.

  On the other hand, as shown in FIG. 3B, the connection relationship between each determination node 151a2 is the ID of the determination node and one or a plurality of connection destinations directly connected to the determination node. It can be represented by a table showing the correspondence with the ID of the determination node. In the table of FIG. 3B, for example, the decision node 1 has a decision node 151 a 2 with ID “3” (hereinafter also referred to as “determination node 3” (hereinafter, the same)) from the decision node 1 to the decision node 3. It shows that it is directly connected in the direction of. In addition, the determination node 2 indicates that no other determination node is connected in the direction from the determination node 2 to the other determination node. In addition, it is shown that two determination nodes, determination node 1 and determination node 8, are directly connected to determination node 4 in the direction from determination node 1 to determination node 1 or determination node 8.

  The configuration of the directed graph 151 can be represented by the two types of tables shown in FIG. In other words, by using not only the ID but also the information indicating the type of the node, it is possible not to distinguish between each start node or each decision node, but to distinguish between each node common to the start node and the decision node. A connection relationship can be expressed. Further, as described above, the nodes in the directed graph 151 are not provided with two types of nodes, i.e., a start node and a determination node. By adding information representing the starting point of the above, it is possible to give a part of the determination node the function of the starting node.

  On the other hand, the determination node content information 151b shown in FIG. 2 can be configured in a data format as shown in FIG. 4, for example. In the example illustrated in FIG. 4, the determination node content information 151b includes a field 151b1 indicating a determination node ID, a field 151b2 indicating attribute information, a field 151b3 indicating a determination expression, and a field 151b4 indicating frequency information. ing. In this example, “frequency information” corresponds to the “predetermined parameter corresponding to the reliability of the determination formula” described above. This “frequency information” is information indicating the frequency of appearance of attribute information associated with the determination node in a plurality of training program information used when the preference model 151 is generated or updated. However, instead of the frequency information, a result of calculating a value representing the reliability of the determination formula by performing a predetermined calculation process on the frequency information may be included in the determination node content information 151b. For example, a value obtained by dividing the value of each frequency by the value of the highest frequency among all the determination nodes (or a value obtained by further weighting the value in a non-linear manner), etc. Parameter ". The “predetermined parameter corresponding to the reliability of the judgment formula” is, for example, a value such that “0” has the lowest reliability and “1” has the highest reliability as real numbers “0” to “1”. Can be set to

  As shown in FIG. 4, the attribute information may be represented by a combination of the type and content of the attribute information, may be represented by a keyword that is not related to the type, or both configurations are mixed. May be. Further, for example, as shown in FIG. 4, the frequency information includes the frequency included in the program information set as “like” and the “dislike” in training when the preference model 151 is generated and updated. It can be divided and included in the frequency included in the set program information. Alternatively, the frequency information may represent the frequency at which the attribute information appears in the program information used in the training regardless of the “like and dislike” setting.

  Further, the determination formula can be in the following format, for example. That is, the judgment formula has the highest probability that the value of the judgment formula (that is, the calculation result) is “-1” to “1”, for example, “−1” is “dislike”, and “1” is “1”. A value indicating that the probability of “like” is the highest can be output. As a variable (that is, parameter) used in the determination formula, for example, frequency information may be used, or a value (that is, a constant) set in advance according to the type and content of attribute information may be used. For example, when frequency information is used as a parameter, the total frequency of “like” and “dislike” is not zero, but when the frequency is set to “like” is zero (that is, all are set to “dislike”) The value of the judgment formula is “−1”. Further, when the frequency when “dislike” is set is zero (that is, when all “like” is set), the value of the determination formula is “1”. Then, when both the “like” frequency and the “dislike” frequency are not zero, the contents of the determination formula are configured so as to be an intermediate value. In this case, the determination formula may be configured using one or a plurality of calculation formulas, or may be configured using a conditional statement (for example, “if” statement) in the program.

  Next, a description will be given of a procedure for generating a logical expression representing a rule for determining whether or not to like a program from the directed graph 151a shown in FIG. In the following description, it is assumed that the attribute information shown in FIG. 5 is associated with the eight determination nodes. That is, “drama” is determined for the determination node 1, “soccer” for the determination node 2, “13:00” for the determination node 3, “broadcast station A” for the determination node 4, and “Movie”, “Actor B” is associated with the decision node 6, “Special number” is associated with the decision node 7, and “News” is associated with the decision node 8. The contents of each determination node have the format shown in FIG. That is, each of the determination nodes 1 to 8 is associated with determination formulas f1 to f8 that output the degree of “like” or “dislike” of the node alone. Further, each of the determination nodes 1 to 8 has a value n1a to n8a ("like") corresponding to the frequency at which the attribute information associated with the determination node appears in the training program information as a parameter indicating reliability. ) And n1b to n8b (in the case of “dislike”) are set.

  When a rule is extracted from the directed graph 151a, as shown in FIG. 7, the determination at each determination node is performed by following the determination nodes connected by directed edges (that is, arrows) from the start nodes 1 to 3, respectively. Combine the judgment results by formula. Each rule can have a start node as a start point and an end point of any one of 0, 1, or a plurality of determination nodes that can be followed in the direction of the arrow. Rules that may be extracted from the directed graph 151a are shown in the balloon 151c. When extracting a rule, the same judgment node is not included twice or more to avoid rule circulation. For example, the rule that follows the determination node 4, the determination node 1, the determination node 3, the determination node 7, and the determination node 1 with the start node 1 indicated by a chain line in FIG. Since it is included twice, it cannot be extracted.

  Next, how an actual logical expression is expressed from the conceptual diagram of the extracted rule will be described with reference to FIG. The logical expression of the extracted rule has an expression form such as “if X1 and X2 and X3 then Y”. At this time, this indicates that the attribute information of the program associated with a certain determination node i is included in the input program information. Y is an instruction obtained from all the determination expressions fi of the determination node i, and is an instruction of “like” or “dislike”. However, the rule reliability parameter, which is a parameter calculated from the “predetermined parameter corresponding to the reliability of the determination formula” of the determination node i, is set in this rule.

  As an example, “broadcasting station A”, “drama”, and “13:00” are connected as one rule, and the result obtained by combining the determination formulas f4, f1, and f3 of the respective determination nodes 4, 1, and 3 is “ Consider a rule with a high probability of “like”. For example, it is assumed that the overall determination of the determination formula is “the probability of“ like ”is higher as the value of (f4 + f1 + f3) is larger”. In this case, the determination formulas f4, f1, and f3 each output a value from “−1” to “1” (here, the closer to “1”, the higher the probability of “like”). For example, it is assumed that the determination formulas f4, f1, and f3 are all values close to “1”. At this time, since the “value of (f4 + f1 + f3)” becomes large (for example, becomes larger than a predetermined threshold value), the rule command becomes ““ like ””. In this case, the rule generated from the start node 1 and the determination nodes 4, 1 and 3 returns a determination result of “like” if a certain program is “Broadcasting station A”, “Drama” and “13:00”. It becomes a logical expression. Further, the rule reliability parameter can be set from the frequency information n4a, n4b, n1a, n1b, n3a, and n3b of the decision nodes 4, 1, and 3, respectively. For example, the reliability parameter of the rule is calculated so that the reliability increases as the integrated value, maximum value, minimum value, or average value of the frequency (or the value obtained by dividing it by the maximum frequency of all judgment nodes) increases. To do.

As another example, “movie” and “actor B” are connected as one rule, and the result of combining the determination expressions f5 and f6 of the respective determination nodes 5 and 6 has a high probability of “like”. Think about the rules. Similar to the above, the overall judgment of the judgment formula is “(f5 + f6)”.
The determination formulas f5 and f6 each output a value from “−1” to “1” as the value of “I like” increases as the value of “is larger”. If the determination formulas f5 and f6 are all close to “1”, for example, (f5 + f6) becomes large (eg, becomes larger than a predetermined threshold value). I like "". In this case, the rule generated from the start node 2 and the determination nodes 5 and 6 is a logical expression that returns a determination result of “like” if a certain program is “movie” and “actor B”. Further, the reliability parameter of this rule can be set from the frequency information n5a, n5b, n6a, and n6b of each determination node 5 and 6.

  Furthermore, as another example, “movie” and “special number” are connected as one rule, and the result of combining the judgment expressions f5 and f7 of the respective judgment nodes 5 and 7 has a high probability of “dislike”. Think about the rules. Similarly to the above, the overall judgment of the judgment formula is set as “the probability of“ like ”is higher as the value of (f5 + f7) is larger”, and the judgment formulas f5 and f7 each output a value from “−1” to “1”. (Here, the closer to “1”, the higher the probability of “likes”, and the closer to “−1”, the higher the probability of “dislikes”). In this case, for example, even if the judgment formula f5 is a value close to “1”, if the judgment formula f7 is a value close to “−1”, (f5 + f7) is not so large (for example, a predetermined value) The rule command has a low probability of “like” (not greater than the threshold). Here, assuming that the case where the probability of “like” is low is “dislike”, the rule generated from the start node 3 and the determination nodes 5 and 7 is “dislike” if a certain program is “movie” and “special number”. This is a logical expression that returns a determination result. Further, the reliability parameter of this rule can be set from the frequency information n5a, n5b, n7a, and n7b of each determination node 5 and 7.

  Next, an exemplary operation of the information processing system 10 illustrated in FIG. 1 will be described with reference to FIGS. FIG. 9 is a flowchart illustrating an example of a processing flow when the information processing system 10 recommends a program that suits the user's preference to the user. FIG. 10 is a flowchart for explaining an example of the processing flow in the content evaluation (step S103) shown in FIG.

  In the information processing system 10, first, the preference model generation unit 150 generates a preference model 151 using the program information for training (step S101). For example, the preference model generation unit 150 appropriately adds a start node and a determination node to the directed graph so that all the attribute information included in the program information for training is included in the preference model 151. Also, determination node content information corresponding to the added determination node is created. There is no particular limitation on the method of generating the directed graph. In the present embodiment, rule selection and the like are automatically performed by a genetic algorithm, which will be described later, that is, a wrinkle process and an evolution process. Therefore, a node can be arbitrarily added to the directed graph.

  Next, the content receiving unit 110 receives program information (step S102). Next, the content evaluation unit 120 and the preference model generation unit 150 cooperate with each other to estimate the user's preference using one or more pieces of program information received by the content reception unit 110 as unknown program information to be evaluated. The content evaluation process is performed (step S103). Details of the content evaluation process will be described later.

  Next, the recommendation result display selection unit 130 displays the program information evaluated as “like” by the content evaluation unit 120 as a recommendation result (step S104). Here, when the user performs a predetermined operation (in the case of “yes” in step S105), the training information setting unit 160 is designated in the program information displayed as the recommendation result according to the user's operation. The program information is set to “like” or “dislike”, and the setting result is temporarily stored in a predetermined storage unit together with the designated program information (step S106).

  Next, when the user selects any content indicated as the recommendation result (in the case of “yes” in step S107), the recommendation result display selection unit 130 receives the content selected by the content reception unit 110. And a display instruction are output (step S108). Next, the program information of the content selected by the recommendation result display selection unit 130 is output to the content browsing log generation unit 140, and the input program information is recorded by the content browsing log generation unit 140 (step S109). .

  Next, the preference model generation unit 150 responds to the new attribute information based on the program information stored by the training information setting unit 160 in step S106 and the program information recorded by the content browsing log generation unit 140 in step S109. The preference model 151 is updated by performing a process of adding a node to the preference model 151 or updating the appearance frequency of an existing determination node (step S110). Thereafter, the processing from step S102 is executed again at a predetermined timing, for example.

  Next, the content evaluation processing in step S103 will be described with reference to FIG. First, as described above with reference to FIGS. 7 and 8, the content evaluation unit 120 sets a set of rules using “like” as a command from the preference model 151 (ie, determines “like”) and “ A set of rules having “dislike” as a command (that is, determining “dislike”) is extracted (step S201). At that time, the content evaluation unit 120 holds information indicating the rule newly generated from the preference model 151 and the combination of nodes used when generating the rule as history information regarding rule generation. Then, when a combination of nodes other than the combination of nodes already recorded in the history information is detected, the content evaluation unit 120 performs a process of extracting a new rule. Therefore, no rule is generated for a combination of nodes that have already been extracted.

  Next, the content evaluation unit 120 executes a wrinkle process on the rule using the training program information (step S202). In the present embodiment, it is assumed that a rule can be generated for all combinations of nodes in the directed graph 151 as shown in FIG. Therefore, if all the generated rules are stored, a large number of rules are held. Therefore, in this embodiment, a rule that is not effective in estimating the preference is hesitant from a rule that continues to be generated in a large amount in the course of evolutionary computation by the following trap process.

  FIG. 11 shows a flow of the wrinkle process using a schematic diagram. In the present embodiment, evolution processing as used in the genetic algorithm is executed on the preference model 151 in step S204 described later. Before and after the evolution process, the preference model 151 can be defined as a model of a different generation. Further, as shown in FIG. 11, the rule set 302 extracted from the preference model 151 in a certain generation is grouped into a rule group 302a that makes a determination of “like” and a rule group 302b that makes a determination of “dislike”. Can be The content evaluation unit 120 holds a certain number of these rules as a set. At this time, the content evaluation unit 120 executes a rule habit process according to the following procedure.

  (Step S301) First, when the program information 301 for training is input, the content evaluation unit 120 applies the program information 301 to each rule, thereby obtaining a determination result of “like” or “dislike” in all the rules. calculate.

  (Step S302) Next, the content evaluation unit 120 statistically processes all the determination results (that is, based on the total value of the determination results, for example), and recommends whether the program is “like” or “dislike”. Finally, 303 is determined. At this time, the degree of influence on the overall determination result can be normalized by multiplying each determination result by the reliability of the rule.

    (Step S303) Next, the content evaluation unit 120 compares the recommendation result 303 and the correct answer value 304 previously given to the program information for training, and determines that the rule that matches is a good rule and the rule that does not match is a bad rule. I will rank.

  By repeating the above operation (steps S301 to S303) for a large number of pieces of training program information, the rules ranked lower are deceived, and only the rules ranked higher are left.

  Then, the content evaluation unit 120 recursively advances the rule selection process (that is, step S202) while performing the generation change process in the evolution process (or evolution calculation) executed in step S204. Only rules that contribute to high recommendation accuracy can be left.

  After executing the selection process using a plurality of training program information in step S202, in step S203, the content evaluation unit 120 determines whether or not the number of newly appearing rules is equal to or less than a predetermined threshold. . Here, a case where a rule newly appears is a phenomenon that occurs when the evolution process in the next step S204 is executed. Further, if the evolutionary calculation and the wrinkle processing sufficiently proceed, new rules that contribute to the recommendation do not appear much even if the wrinkle processing as shown in FIG. 11 is performed. In this system, when the number of newly appearing rules falls below a predetermined threshold value, it is determined that the learning process has ended, and the evolution calculation ends. According to this configuration, in the estimation process in step S205 described later, unknown program information is input, and the user's preference is determined based on the learned rule group. Therefore, high accuracy is expected.

  In the evolution calculation using the genetic algorithm of this embodiment in step S204, crossover processing and mutation processing are performed in the same manner as a general genetic algorithm. These crossover processing and mutation processing can be performed by the preference model generation unit 150. Hereinafter, these two processes will be described.

  As shown in FIG. 12, in the crossover process of the present embodiment, the same determination nodes holding connection information to the next node are exchanged for two effective graphs (that is, genes in the genetic algorithm) serving as parents. . As a result, the new effective graph (gene) has a new graph structure that combines the directed graphs of the two genes.

  Further, as shown in FIGS. 13A and 13B, in the mutation process of the present embodiment, the determination node number is changed at random or the configuration information of the graph representing the next connection destination (FIG. 3 ( b)) is changed randomly.

  In steps S201 to S204, it is possible to leave only a rule that contributes to high recommendation accuracy by repeatedly extracting a set of rules and repeatedly executing the selection process and the evolution process. When the number of newly appearing rules is below a predetermined threshold, the determination result in step S203 is “no”.

  Next, in step S205, the content evaluation unit 120 estimates the user's preference by applying each rule generated based on the preference model 151 to the input program information. That is, as shown in FIG. 14, the content evaluation unit 120 determines that the unknown program information 401 is determined to be “dislike” and the rule group 402 a that is determined as “like” selected in steps S <b> 201 to S <b> 204. The rule 402 is applied to each rule of the rule set 402 including the group 402b (step S401). Next, the content evaluation unit 120 statistically processes the determination results of each rule, determines likes and dislikes, and determines a recommendation result 403 (that is, information indicating whether or not the program information should be recommended) (step S402). ).

  Next, in step S206 of FIG. 10, the content evaluation unit 120 determines whether or not evaluation for all program information for which content evaluation processing should be executed has been completed. If other program information to be evaluated remains, the process of step S205 is executed again for the next program information (from “no” in step S206 to step S205). If there is no remaining program information to be evaluated, the process ends (in the case of “yes” in step S206).

  As described above, according to the present embodiment, there is no need for a parent-child relationship between information associated with each determination node and information associated with another determination node. Therefore, it is possible to easily add or delete determination nodes from the preference model. In addition, since the information for generating or updating the preference model according to the present embodiment can be acquired from content information such as program information, an increase in traffic volume and restrictions on use in an offline environment Such a problem does not occur.

  Further, according to the present embodiment, the user's preference regarding the contents can be expressed by a set of mathematical expressions, and the preference can be determined by giving unknown contents to these mathematical expression sets. In other words, the user's preference regarding the content can be modeled as a mathematical formula, so that the result is highly visible.

  In addition, since the content recommendation system of this embodiment can automatically extract information necessary for expressing user preferences from a large amount of input information and create rules, it can be expected to have good recommendation accuracy even for large-scale content data. Is done. In addition, since the content recommendation system using the genetic algorithm can operate even in a completely offline environment, it is also useful for devices that cannot be connected to an external environment such as a network.

  The features of the present invention can be summarized as follows, for example. It can be said that the present invention uses one of calculation methods for automatically generating a program by connecting nodes called GNP (Genetic Network Programming) in a network. Here, Feature 1 and Feature 2 below are unique contrivances when applying the GNP algorithm to the recommendation system, and Feature 3 is a difference from the technology behind the recommendation system itself.

Feature 1: The reliability information is entered in the decision node in each calculation in the GNP algorithm.
The evolution calculation by GNP has a disadvantage that the calculation cost is enormous in addition to the advantage that a complicated rule can be calculated. Therefore, in order to apply the GNP algorithm to the recommendation system, a method for efficiently searching for significant rules from a large number of decision nodes is important. In the present invention, by introducing reliability information, a determination node including attribute information frequently appearing in training data (that is, training program information) is devised so as to be easily adopted as a part of the rule. . With the introduction of this reliability, the convergence time of the rule is greatly increased, and the GNP algorithm can be driven in real time even in the recommendation system.

Feature 2: Even in the recommendation process, the degree of influence of the rule is normalized by using the reliability information of the judgment node.
When the recommendation process is performed using the rule group adopted by the system of the present invention, a method is conceivable in which the determination results based on all the rules are totaled and determined by majority decision of the “like” and “dislike” determination results. In fact, the proposed system also uses this method. However, depending on the rules, there are some that have been successfully discriminated accidentally even though they are not related to the training data information. Such a rule may cause the determination process to fail when recommending unknown program information. In order to avoid this problem, the reliability information introduced in Feature 1 has been devised so that it can be used in the “like” and “dislike” discrimination processing using rules. The reliability information represents the appearance frequency of attribute information in the training data. Therefore, the reliability of the rule calculated based on the reliability represents how correlated with the training data. In the recommendation process, a higher recommendation accuracy was confirmed by normalizing the influence of a rule with the reliability of the rule.

Feature 3: The recommendation system has high resolution of preference (likes and dislikes).
The following describes the advantages of the recommendation system itself using GNP, unlike other recommendation systems. Many of the algorithms used in other recommendation systems are two-class classifiers that can only be classified as “like” or “dislike”. In contrast, this method of outputting the determination result by a logical expression can be classified into N classes. Actually, the system of the present invention can output the results in a plurality of categories such as “I like it very much”, “I like it a lot”, “I like it so much”, and “I like it so much”. Therefore, depending on the request of the application, an unexpected recommendation result such as “somewhat like” can be output in addition to “very like”.

DESCRIPTION OF SYMBOLS 10 Information processing apparatus 100 Content display apparatus 120 Content evaluation part 130 Recommendation result display selection part 140 Content browsing log production | generation part 150 Preference model production | generation part 151 Preference model 151a Directed graph 151a1 Start node 151a2 Judgment node 151a3 Edge 151b Judgment node content information 200 Content control Device 210 Content control unit

Claims (7)

Represented by a directed graph composed of a plurality of nodes and a plurality of edges connecting the plurality of nodes with directionality,
The plurality of nodes are start nodes or decision nodes;
The determination node determines one attribute information of the plurality of attribute information included in the content information representing the attribute of the content using the plurality of attribute information, and a user's preference degree related to the determination node. Is associated with information representing a judgment formula for
The start node makes a predetermined determination regarding user preference based on one or more determination results based on one or more of the determination formulas associated with one or more of the determination nodes following the start node. A starting point of a program for generating one or a plurality of rules for connection with connection information to one or a plurality of the determination nodes,
A storage unit storing a preference model having a data structure;
An information processing apparatus comprising: an estimation unit that inputs the content information and estimates one's preference by applying one or more rules generated based on the preference model to the content information. The determination node is further associated with a predetermined parameter for representing the reliability of the determination formula associated with the determination node.
The rule generated based on the preference model includes the one or more representing the reliability of the one or more determination formulas obtained from the one or more determination results used as a reference when generating the rule Is a value obtained using a predetermined parameter, and a value representing the reliability of the rule is set,
The said estimation part utilizes the value showing each said reliability set to each said rule, when applying the said 1 or several rule with respect to the said content information. Information processing device. The preference model is generated or updated based on a plurality of the content information,
The attribute information associated with the determination node appears in the plurality of content information used as a reference when the predetermined parameter associated with the determination node generates or updates the preference model. The information processing apparatus according to claim 2, wherein the information processing apparatus corresponds to a frequency. The information according to any one of claims 1 to 3, wherein the estimation unit estimates the user's preference while performing an evolution process on the preference model using a genetic algorithm. Processing equipment. Represented by a directed graph composed of a plurality of nodes and a plurality of edges connecting the plurality of nodes with directionality,
The plurality of nodes are start nodes or decision nodes;
The determination node determines one attribute information of the plurality of attribute information included in the content information representing the attribute of the content using the plurality of attribute information, and a user's preference degree related to the determination node. Is associated with information representing a judgment formula for
The start node makes a predetermined determination regarding user preference based on one or more determination results based on one or more of the determination formulas associated with one or more of the determination nodes following the start node. A starting point of a program for generating one or a plurality of rules for connection with connection information to one or a plurality of the determination nodes,
Storing a preference model having a data structure in a predetermined storage unit;
Information processing characterized by inputting the content information by an estimation unit and estimating the user's preference by applying one or more rules generated based on the preference model to the content information Method. A data structure for representing a user preference model,
Represented by a directed graph composed of a plurality of nodes and a plurality of edges connecting the plurality of nodes with directionality,
The plurality of nodes are start nodes or decision nodes;
The determination node determines one attribute information of the plurality of attribute information included in the content information representing the attribute of the content using the plurality of attribute information, and a user's preference degree related to the determination node. Is associated with information representing a judgment formula for
The start node makes a predetermined determination regarding user preference based on one or more determination results based on one or more of the determination formulas associated with one or more of the determination nodes following the start node. A starting point of a program for generating one or a plurality of rules for connection with connection information to one or a plurality of the determination nodes,
A data structure characterized by that. The data structure according to claim 6, wherein the determination node is further associated with a predetermined parameter for representing the reliability of the determination formula associated with the determination node.

Images