JP2018005918A - Virtual evaluation system and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To predict feature quantity for an unknown item for a user from the feature quantity determined by other user using correlation of feature quantity perception among the users.SOLUTION: An input-output relation in which feature quantity of an item determined by an expert is output for an input of the feature quantity of the item determined by a user for each combination of the user and the expert on the basis of user item information and expert item information including the feature quantity and an evaluation value of the item for each user and for each expert, the input-output relation in which the feature quantity of the item determined by the user is output for the input of the feature quantity of the item determined by the expert, and the input-output relation in which the evaluation value of the item is output for the input of the feature quantity of the item by the user are learned. Then, a predicted evaluation value for the item of the user is predicted based on the user item information, the expert item information, and a learning value.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a virtual evaluation system and a program.

  Consider a machine learner that outputs a predicted evaluation value for an item when a feature value of the item is input. In general, in such a machine learning device, it is required to calculate the feature amount as accurately as possible for as many items as possible in advance (see, for example, Non-Patent Document 1).

AISSY Corporation Company / Technology Overview, http://aissy.co.jp/images/AISSY-1218.pdf, Dec. 18th, 2014

  However, when the feature amount determination cost is high, for example, when a human professional checks the item and determines the feature amount, or when using a sensor that requires a large cost in terms of time and money for measurement. However, there is a problem that the feature quantity of an accurate and large amount of items cannot be obtained.

  The virtual evaluation system of the present invention solves the above-mentioned problem by predicting a feature quantity for an item unknown to a user from a feature quantity determined by another user by using a correlation of feature quantity perception between users. The main purpose is to present the means to do this. The main purpose of the program of the present invention is to make a computer function as the virtual evaluation system.

  The virtual evaluation system of the present invention employs the following means in order to achieve the main object described above.

The virtual evaluation system of the present invention is
A virtual evaluation system that specifies and recommends items to be recommended when an item recommendation request is made by a user,
User item information including the feature amount and evaluation value of the item determined by the user for each item for each user, and expert item information including the item feature amount and the evaluation value determined for each item by the expert for each expert And a storage unit for storing information including:
Based on the user item information and the expert item information, for each combination of the user and the expert, a user who outputs the feature amount of the item determined by the expert in response to the input of the item feature amount determined by the user For the expert input / output relationship, the expert user input / output relationship for outputting the feature amount of the item determined by the user in response to the input of the feature amount of the item determined by the expert, and the input of the feature amount of the item by the user A user item evaluation learning unit that learns a user item evaluation input / output relationship that outputs an evaluation value of the item; a learning unit that learns;
A predicted evaluation value prediction unit that predicts a predicted evaluation value for a user item based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship;
A recommendation processing unit that recommends an item based on the predicted evaluation value when an item recommendation request is made from a user;
It is a summary to provide.

  In this virtual evaluation system of the present invention, user item information including the item feature amount and evaluation value determined by the user for each item for each user, and item feature amount determined by the expert for each item for each expert And the expert item information including the evaluation value, and the feature amount of the item determined by the user for each combination of the user and the expert based on the user item information and the expert item information. User / expert input / output relationship that outputs the feature value of the item determined by the expert in response to the input, and the expert user that outputs the feature value of the item determined by the user in response to the input of the feature value of the item determined by the expert Input / output relations and item feature input by user A user item evaluation learning unit that learns the user item evaluation output relationship for outputting an evaluation value of the item with respect to learning. A predicted evaluation value for a user item is predicted based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship. Thereby, the feature quantity with respect to the item unknown to the user can be predicted from the feature quantity determined by another user by using the correlation (input / output relation) of the feature quantity perception between the user and the expert.

  In such a virtual evaluation system of the present invention, the predicted evaluation value predicting unit is characterized by at least one item feature amount for each user based on the user item information, the expert item information, and the user / expert input / output relationship. First, the feature amount of the item perceived by the expert is predicted for all experts that have not been determined, and the feature amount of all items in the expert is determined based on the expert item information and the predicted feature amount. Based on the item feature amount determination unit, the expert user input / output relationship and the feature amount of all items determined by the first item feature amount determination unit, predicting the feature amount perception of the user item, All items in the user using feature perception Predicting the predicted evaluation value based on a second item feature amount determination unit that determines a feature amount, the user item evaluation input / output relationship, and the feature amounts of all items determined by the second item feature amount determination unit It is also possible to have a predicting unit to perform. In this way, a more appropriate prediction evaluation value can be predicted.

  Further, in the virtual evaluation system of the present invention, the predicted evaluation value predicting unit is configured so that the expert for each user and each expert based on the user item information, the expert item information, and the user / expert input / output relationship. A first item feature amount determination unit that predicts feature amounts of items to be perceived and determines feature amounts of all items in the expert based on the expert item information and the predicted feature amounts; and the expert user input / output Based on the relationship and the feature amount of all items determined by the first item feature amount determination unit, the feature amount perception of the user's item is predicted, and the feature amount of all items in the user using the feature amount perception A second item feature amount determining unit for determining the user item, and the user item Based on the feature quantity of all items that have been determined rated output relationship between the second item characteristic amount determination unit, may be assumed to have, a prediction unit that predicts the predictive evaluation value. In this way, a more appropriate prediction evaluation value can be predicted.

  In these aspects of the virtual evaluation system of the present invention, the first item feature amount determination unit determines feature values of all items in the expert by obtaining an average value of the predicted feature amounts for each item, The two-item feature amount determination unit may determine the feature amounts of all items in the user by obtaining an average value of the predicted feature amounts for each item.

  The virtual evaluation system of the present invention may include a recommendation processing unit that recommends an item based on the predicted evaluation value when an item recommendation request is made from a user. For example, an item with the highest predicted evaluation value is recommended, or a plurality of items are recommended in order of the highest predicted evaluation value. Using the correlation (input / output relationship) of the feature amount perception between the user and the expert, the feature amount for the item unknown to the user is predicted and recommended from the feature amount determined by other users. You can recommend items that suit your taste.

The program of the present invention
A program for causing a computer to function as a virtual evaluation system,
User item information including the feature amount and evaluation value of the item determined by the user for each item for each user, and expert item information including the item feature amount and the evaluation value determined for each item by the expert for each expert And memorize information including
Based on the user item information and the expert item information, for each combination of the user and the expert, a user who outputs the feature amount of the item determined by the expert in response to the input of the item feature amount determined by the user For the expert input / output relationship, the expert user input / output relationship for outputting the feature amount of the item determined by the user in response to the input of the feature amount of the item determined by the expert, and the input of the feature amount of the item by the user Learning a user item evaluation learning unit that learns a user item evaluation input / output relationship that outputs an evaluation value of the item;
Predicting a predicted evaluation value for a user's item based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship;
It is summarized as having.

  The program of the present invention includes user item information including item feature amounts and evaluation values determined by the user for each item for each user, and item feature amounts and evaluation values determined by the expert for each item for each expert. For each item of the combination of the user and the expert based on the user item information and the expert item information. The user-expert input / output relationship that outputs the feature value of the item determined by the expert and the expert-user input / output relationship that outputs the feature value of the item determined by the user in response to the input of the item feature value determined by the expert And the item against the input of the feature amount of the item by the user A user item evaluation learning unit that learns the user item evaluation output relationship for outputting an evaluation value, the learned. A predicted evaluation value for a user item is predicted based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship. As a result, the computer can function as the virtual evaluation system of the present invention. Therefore, the computer functioning as the virtual evaluation system uses the correlation (input / output relationship) of the feature amount perception between the user and the expert, and the feature amount for the item unknown to the user is determined by the feature determined by another user. Can be predicted from the quantity.

It is a block diagram which shows the outline of a structure of the virtual evaluation system 20 as one Example of this invention. It is explanatory drawing which shows an example of the flow of the item evaluation value prediction in the virtual evaluation system 20 of an Example. It is a flowchart which shows an example of a user's item evaluation. It is a flowchart which shows an example of item evaluation of an expert. It is a flowchart which shows an example of the learning by the U-> E feature-value calculator. 5 is a flowchart showing an example of learning by an E → U feature quantity calculator 26; It is a flowchart which shows an example of learning by the user item evaluation estimated value calculator. It is a flowchart which shows an example of prediction.

  Next, the form for implementing this invention is demonstrated using an Example.

  FIG. 1 is a configuration diagram showing an outline of a configuration of a virtual evaluation system 20 as an embodiment of the present invention. As shown in FIG. 1, the virtual evaluation system 20 of the present invention is configured to be connectable to a user terminal 12 and an expert terminal 14, and includes a Web server 22, a user expert feature quantity calculator (U → E feature quantity calculator). 24, expert user feature quantity calculator (E → U feature quantity calculator) 26, user item evaluation predicted value calculator 28, item information table 30, user item evaluation table 32, expert item evaluation table 34, user / expert learner table ( U → E learner table) 36, expert user learner table (E → U learner table) 38, user item evaluation predictor table 40, and user item evaluation prediction value table 42. The U → E feature quantity calculator 24 is a calculator equipped with a learning device for predicting expert feature quantity perception from the user's feature quantity perception, and the U → E learner table 36 is the U → E feature quantity calculator 24. It is a table for preserve | saving the parameter for comprising this learning device. The E → U feature value calculator 26 is a computer including a learning device for predicting the user's feature value perception from the expert feature value perception, and the E → U learner table 38 is the E → U feature value calculator 26. It is a table for preserve | saving the parameter for comprising this learning device. In FIG. 1, the Web server 22, the U → E feature quantity calculator 24, the E → U feature quantity calculator 26, and the user item evaluation predicted value calculator 28 are configured by individual computers, but a plurality of functions that also serve as a part of each function are provided. A computer or a single computer may be used. Item information table 30, user item evaluation table 32, expert item evaluation table 34, U → E learner table 36, E → U learner table 38, user item evaluation predictor table 40, user item evaluation prediction value table 42 A plurality of storage devices may be configured, or a single storage device may be configured.

A. Operation of Virtual Evaluation System 20 Next, the operation of the virtual evaluation system 20 of the embodiment will be described. FIG. 2 is an explanatory diagram illustrating an example of a flow of item evaluation value prediction in the virtual evaluation system 20 of the embodiment.

(1) Prior Item Evaluation Assume that the number set of I items is I = {1,..., I}, and the number set of U users is U = {1,. Separately from the user, a number set of E experts (which may be considered as sensors) that can accurately determine the feature amount of an item is E = {1,..., E}. A flowchart showing an example of user item evaluation is shown in FIG. 3, and a flowchart showing an example of expert item evaluation is shown in FIG.

  As shown in FIG. 3, the user u (formula (1)) accesses the web server 22 from the user terminal 12 and sets the number (identifier) i (formula (2)) of the item to be evaluated to the web server 22. Send and request information presentation of the item (step S100). The Web server 22 reads information on the item from the item information table 30 (Step S110) and presents it to the user (Steps S120 and S130). The user u determines the feature amount x (formula (3)) and the evaluation value d (formula (4)) of the item, and sends the item number (identifier) i and the user number (identifier) u to the Web server 22. Send (step S140). The Web server 22 stores the sent user number, item number, feature amount, and evaluation value in the user item evaluation table 32 (step S150).

  As shown in FIG. 4, the expert e (formula (5)) accesses the web server 22 from the expert terminal 14, and sends the number (identifier) i (formula (2)) of the item to be evaluated to the web server 22. Send and request information presentation of the item (step S200). The Web server 22 reads the item information from the item information table 30 (step S210) and presents it to the expert (steps S220 and S230). The expert e determines the feature amount (formula (6)) of the item and sends it to the Web server 22 together with the item number (identifier) i and the expert number (identifier) e (step S240). The Web server 22 stores the sent expert number, item number, and feature quantity in the expert item evaluation table 34 (step S250).

(2) Learning In the above-described prior item evaluation, the item number set for which the user u (formula (1)) has determined the feature quantity is represented by formula (7), and the feature quantities are represented by formula (3) and formula (8). The evaluation values are represented by Equation (4) and Equation (8). The item number set for which the expert e (formula (5)) has determined the feature quantity is represented by formula (9), and the feature quantities are represented by formula (10) and formula (11). In the following description, | A | represents the number of elements of the set A. A flowchart showing an example of learning by the U → E feature quantity calculator 24 is shown in FIG. 5, a flowchart showing an example of learning by the E → U feature quantity calculator 26 is shown in FIG. 6, and learning by the user item evaluation predicted value calculator 28 is performed. A flowchart showing an example is shown in FIG.

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34 (step S300), and the user u (formula (1)) and the expert are related to the portion indicated by A in FIG. For all combinations of e (formula (5)), the correlation between the feature amount perception of the user and the expert is learned using an arbitrary machine learning device (step S310). A set of item numbers for which feature quantities are determined by both the user u and the expert e is represented by Expression (12). In this correlation learning, when the feature amount of the item determined by the user u ((Expression (13), (Expression (14))) is input, the feature amount of the same item determined by the expert e (Expression (15)), the expression (14)) is output so that the input / output relationship (formula (16)) is learned.The U → E feature quantity calculator 24 uses the user number u, the expert number e, and the parameters (learning results) necessary for the learning device configuration. ) Is stored in the U → E learner table 36 (step S320).

Procedure (2)
The E → U feature quantity calculator 26 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34 (step S400), and the user u (formula (1)) and the expert for the part indicated by B in FIG. For all combinations of e (formula (5)), the inverse function of the above input / output relationship (formula (16)) is learned using an arbitrary machine learning device (step S410). That is, when the feature amount ((expression (15)), expression (14)) determined by the expert e is input, the feature amount (expression (13), expression (14)) determined by the user u is output. Thus, the input / output relationship (formula (17)) is learned, and the E → U feature quantity calculator 26 performs E → U learning on the user number u, expert number e, and parameters (learning results) necessary for the learning device configuration. The data is stored in the container table 38 (step S420).

Procedure (3)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation table 32 (step S500), and the user u perceives all the users u (formula (1)) with respect to the portion indicated by C in FIG. When the feature amount (formula (13)) of the item i (formula (8)) is input, an input / output relationship (an arbitrary machine learning device is used to output the evaluation value (formula (18)). Equation (19) is learned (step S510). The user item evaluation prediction value calculator 28 stores the user number u and parameters (learning results) necessary for the learning device configuration in the user item evaluation prediction device table 40 (step S520). Any machine learning device may be used as long as it satisfies the input / output requirements.

(3) Prediction Next, a procedure for predicting an evaluation value of an item i (expression (21)) unknown to the user u ′ (expression (20)) using the virtual evaluation system 20 of the embodiment will be described. It is assumed that the feature amount of the item i is determined by any user, that is, Equation (22) holds. Further, among all the experts, the feature amount of the item i is determined, that is, a set of experts e (formula (5)) which is formula (11) is defined as formula (23). Assume that a set of users u (formula (1)) that determines the feature amount of item i among all users, that is, formula (8) is formula (24). A flowchart showing an example of the prediction is shown in FIG. In the flowchart of FIG. 8, the processes of steps S600 to S630 are executed by the U → E feature quantity calculator 24, the processes of steps S640 to S660 are executed by the E → U feature quantity calculator 26, and the processes of steps S670 to S690 are performed by the user. This is executed by the item evaluation predicted value calculator 28.

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32, the expert item evaluation table 34, and the U → E learner table 36 (step S600), and for each user u (formula (25)), For all experts e (formula (26)) for which the feature quantity of item i has not been determined, the feature quantity perceived by expert e is predicted as formula (27) (step S610).

Procedure (2)
The U → E feature quantity calculator 24 reads necessary information from the expert item evaluation table 34 (step S620), and uses the expressions (28), (25), and (26) calculated in the above procedure (1). , The feature amount (formula (29)) of the item i is determined for all experts e (formula (5)). The expert who has already determined the feature amount of the item i, that is, the expert that is the expression (30), is represented by the expression (31). For the expert represented by equation (26), equation (33) is constructed using equations (32) and (25). As a configuration method, for example, a method using an average value of Expressions (28) and (25), that is, a method of Expression (34) is conceivable. Further, the configuration method may be learned in advance using information in which all users and all experts determine the feature amount of the same item, and the result may be used.

Procedure (3)
The E → U feature quantity calculator 26 reads necessary information from the E → U learner table 38 (step S640), and uses the feature quantities (expressions (33) and (5)) calculated in the above procedure (2). The feature amount perception of the item i by the user u ′ is predicted as Equation (35) (step S650).

Procedure (4)
The E → U feature quantity calculator 26 integrates the user's feature quantity perception prediction results using the feature quantity perception (Equation (36), Equation (5)) calculated in the above step (3), and the feature amount perception ( Formula (37)) is determined (step S660). As a determination method, for example, a method using an average value of Equations (36) and (5), that is, a method of Equation (38) can be considered. Further, the configuration method may be learned in advance using information in which all users and all experts determine the feature amount of the same item, and the result may be used.

Procedure (5)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation predictor table 40 (step S670), and uses the equation (37) calculated in the above procedure (4) for the item i of the user u ′. The evaluation value is predicted as equation (39) (step S680). The user item evaluation predicted value calculator stores the user number u ′, the item number i, and the predicted evaluation value (the left side of Expression (39)) in the user item evaluation predicted value table 42 (step S690).

  In the above-described processing, the feature amount of the item i is determined, that is, for the expert who is the equation (30), the method using the determined feature amount as it is is adopted, but the user u (formula (25 A method of correcting the determined feature amount using the feature amount perception information in ()) is also conceivable. When this method is adopted, the above-described procedure (1) and procedure (2) can be changed to the following procedure.

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32, the expert item evaluation table 34, and the U → E learner table 36, and all experts e for each user u (formula (25)). For (Expression (5)), the feature quantity perceived by the expert e is predicted as Expression (27).

Procedure (2)
Subsequently, the U → E feature amount calculator 24 reads necessary information from the expert item evaluation table 34, and uses the equations (28), (25), and (5) calculated in the above procedure (1), The feature amount (formula (29)) of the item i is determined for all experts e (formula (5)). As a configuration method, for example, a method using an average value of Expressions (28) and (25), that is, a method of Expression (40) can be considered. Further, for a user who has already determined the feature amount of the item i, that is, a user who is the equation (30), a method of considering the information of the equation (33) to obtain the equation (41) is also conceivable.

  In the above-described system, the method of handling feature quantity determination by a plurality of experts has been described independently. However, only a single expert is considered, and the feature quantity may be configured by an average value of a plurality of experts. I do not care.

(4) Item recommendation The Web server 22 reads necessary information from the item information table 30 and the user item evaluation predicted value table 42 in response to a request from the user terminal 12, and sends an item to the user based on the item predicted value. To recommend.

B. Specific Example B (1) Wine Recommendation System in which Sommelier Determines Feature Amount Next, a wine recommendation system in which a sommelier determines a feature amount will be described as a specific example using the virtual evaluation system 20 of the present embodiment. In the wine recommendation system, the feature value of wine is expressed in five levels of five variables of sweetness, sourness, bitterness, astringency, and reverberation, and the user's evaluation z (expression (43)) for the wine feature value x (expression (42)). ). The characteristic amount of wine should be evaluated by the sommelier, but the number of sommeliers is limited, and the number of wines that can be evaluated by the sommelier is also limited. Therefore, using the correlation between the feature amount perception of the general user and the sommelier, the feature amount of the wine not evaluated by the sommelier is virtually calculated, and the feature amount perception of the wine of the recommendation target user from this information is calculated. And the evaluation of the wine for the recommendation target user is predicted from this information.

(1) Prior Wine Evaluation Let I = {1,..., I} be the number set of I wines handled by this system, and U = {1,..., U} be the number set of U users. Let E = {1,..., E} be the number set of E sommeliers. The user u (formula (44)) accesses the web server 22 from the user terminal, sends the wine number i (formula (45)) to be evaluated to the web server 22, and requests information presentation of the wine. The Web server 22 reads the wine information from the item information table 30 and presents it to the user. The user u determines the feature amount x (formula (46)) and the evaluation value (formula (47)) of the wine, and sends them to the web server 22 together with the wine number i and the user number u. The Web server 22 stores the sent user number, wine number, feature amount, and evaluation value in the user item evaluation table 32.

  The sommelier e (formula (48)) accesses the web server 22 from the expert terminal 14, sends the number i of the wine to be evaluated (formula (49)) to the web server 22, and requests information presentation of the wine. . The Web server 22 reads the wine information from the item information table 30 and presents it to the sommelier. The sommelier e determines the characteristic amount y (formula (50)) of the wine and sends it to the Web server 22 together with the wine number i and the sommelier number e. The Web server 22 stores the sent sommelier number, wine number, and feature amount in the expert item evaluation table 34.

(2) Learning The wine number set for which the user u (formula (44)) has determined the feature quantity is represented by formula (51), the feature quantity is represented by formula (46), formula (52), and the evaluation is formula (47), Formula (52) is assumed. The wine number set for which the sommelier e (formula (48)) has determined the feature quantity is represented by formula (53), and the feature quantities are represented by formula (50) and formula (54).

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34, and features for all combinations of the user u (formula (44)) and the sommelier e (formula (48)). The correlation of quantity perception is learned, and the wine number set for which both the user u and the expert e have determined the feature amount is represented by the equation (55) The wine feature amount (the equation (56), When the expression (57) is input, the input / output relationship (expression (60)) is set to the sigmoid function so that the characteristic amount (expression (58), expression (59)) of the same wine determined by the sommelier e is output. Learning is performed using a hierarchical neural network as an activation function, and the U → E feature quantity calculator 24 uses a user number u, an expert number e, and an intermediate layer as learner parameters. Stores hidden neurons number, weight between the neurons, the threshold for activation functions U → E learner table 36.

Procedure (2)
The E → U feature quantity calculator 26 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34, and for all combinations of the user u (formula (44)) and the sommelier e (formula (48)), The inverse function of the above equation (60) is learned. In other words, when the wine feature amount ((Expression (58), Equation (59)) determined by the sommelier e is input, the same wine feature amount (Expression (56), Equation (57)) determined by the user u is output. As described above, the input / output relationship (formula (61)) is learned using a hierarchical neural network in which the sigmoid function is an activation function, and the E → U feature quantity calculator 26 has a user number u, an expert number e, As the learner parameters, the number of intermediate layers, the number of intermediate layer neurons, the weight between neurons, and the threshold of the activation function are stored in the E → U learner table 38.

Procedure (3)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation table 32, and for all the users u (formula (44)), the feature quantity (wine (52)) of the wine i perceived by the user u ( When the expression (56)) is input, the input / output relationship (expression (63)) is learned using the radial basis function network so that the evaluation (expression (62)) is output. The user item evaluation predicted value calculator 28 stores the user number u and the number of intermediate layer neurons, the center of the basis function, the radius of the basis function, and the weight between the basis function and the output layer as parameters of the learner in the user item evaluation predictor table 40. save.

(3) Prediction Next, a procedure for predicting the evaluation of wine i (formula (65)) unknown to the user u ′ (formula (64)) will be described using this system. Note that the feature amount of wine i is determined by any user, and the set of user numbers is represented by equation (66). Among all the sommeliers, the number set of the sommeliers that determine the feature amount of the item is represented by Expression (67).

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32, the expert item evaluation table 34, and the U → E learner table 36, and the feature of the wine i for each user u (formula (68)). For all sommeliers e (formula (69)) for which the amount has not been determined, the feature quantity perceived by the sommelier e is predicted as formula (70).

Procedure (2)
The U → E feature quantity calculator 24 reads necessary information from the expert item evaluation table 34 and uses all of the sommeliers using the formulas (71), (68), and (69) calculated in the above procedure (1). For e (formula (48)), the feature quantity (formula (72)) of wine i is determined. The sommelier e (formula (73)) that has already determined the characteristic amount of wine i is represented by formula (74). The sommelier e (formula (69)) for which the feature amount of the wine i has not been determined is represented by formula (75) using formulas (71) and (68) predicted from the user's perceptual information.

Procedure (3)
The E → U feature quantity calculator 26 reads necessary information from the E → U learner table 38, and uses the feature quantities (Equation (27) and Equation (48)) calculated in the above procedure (2), and the user u The feature quantity perception for 'wine i' is predicted as equation (76).

Procedure (4)
The E → U feature value calculator 26 determines the feature value perception of the general user u ′ with respect to the wine i as the equation (78) using the equations (77) and (48) calculated in the procedure (3).

Procedure (5)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation predictor table 40 and uses the equation (79) calculated in the above procedure (4) to evaluate the evaluation of the user u ′ for the wine i by the equation ( 80). The user item evaluation predicted value calculator 28 stores the user number u ′, the wine number i, and the predicted evaluation value (formula (81)) in the user item evaluation predicted value table 42.

(4) Wine recommendation The web server 22 reads necessary information from the item information table 30 and the user item evaluation predicted value table 42 in response to a request from the user terminal 12, and informs the user of wine based on the predicted wine value. To recommend.

B (2) Wine recommendation system in which sensor determines feature quantity Next, a technique for sensing the taste of food such as wine will be considered. Although it is possible to quantitatively measure the taste of food, it is difficult to measure all the wines on the market because of the time and financial cost of measurement. Therefore, we consider benchmark wines that are representative wine collections, perform benchmark sensing in advance, and use the correlation coefficient of feature amount perception by general users and sensors, and feature values of wine not evaluated by the sensor. , The feature amount perception of the recommendation target user for the wine is predicted from this information, and the evaluation of the wine for the recommendation target user is predicted from this information.

(1) Prior Wine Evaluation Let I = {1,..., I} be the number set of I wines handled by this system, and U = {1,..., U} be the number set of U users. The sensed benchmark wine number set is defined as equation (81A), and the feature values are defined as equations (82) and (83). It is assumed that the feature values of the benchmark wine that has been sensed have already been input to the expert item evaluation table 34.

  The user u (formula (84)) accesses the Web server 22 from the user terminal. When the benchmark wine information is not input, the Web server 22 reads the benchmark wine information from the item information table 30 and prompts the user to input the evaluation value. The user u determines the characteristic amount (expression (84), expression (85)) of the benchmark wine and the evaluation value expression ((86), expression (85)), and the web server together with the wine number i and the user number u. 22 to send. The Web server 22 stores the sent user number, wine number, feature amount, and evaluation value in the user item evaluation table 32, and sets the user's benchmark wine information input flag as input. Note that the user does not have to enter information for all benchmark wines. If benchmark wine information has already been input, the number i (formula (87)) of the wine to be evaluated is sent to the Web server 22 to request information presentation of the wine. The Web server 22 reads the wine information from the item information table 30 and presents it to the user. The user u determines the characteristic amount (formula (85)) and the evaluation value (formula (86)) of the wine, and sends them together with the wine number i and the user number u to the Web server 22. The Web server 22 stores the sent user number, wine number, feature amount, and evaluation value in the user item evaluation table 32.

(2) Learning The wine number set expression for which the user u (formula (88)) has determined the feature quantity is (89), the feature quantity is formula (85), formula (90), and the evaluation is formula (86), Let it be Formula (90).

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34, and learns the correlation of feature quantity perception with the sensor for all users u (formula (91)). To do. Of the benchmark wines, the wine number set for which the user u has determined the feature amount is represented by equation (92). When the feature value of the wine determined by the user u (Equation (93), Equation (94)) is input, the input / output relationship ( Equation (96)) is learned using a hierarchical neural network with a sigmoid function as an activation function. The U → E feature quantity calculator 24 stores the user number i, the number of intermediate layers, the number of intermediate layer neurons, the weight between neurons, and the threshold of the activation function as parameters of the learner in the U → E learner table 36.

Procedure (2)
The E → U feature quantity calculator 26 reads necessary information from the user item evaluation table 32 and the expert item evaluation table 34, and learns the inverse function of the above expression (96) for all users u (expression (91)). To do. That is, when the feature value of the benchmark wine (Equation (95), Equation (94)) is input, the same wine feature value (Equation (93), Equation (94)) determined by the user u is output. The output relation (Equation (97)) is learned using a hierarchical neural network with a sigmoid function as an activation function. The E → U feature quantity calculator 26 stores the number of intermediate layers, the number of intermediate layer neurons, the weight between neurons, and the threshold of the activation function in the E → U learner table 38 as the user number u and learner parameters.

Procedure (3)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation table 32 and, for all users u (expression (91)), features (expression (90)) of wine perceived by the user u (expression (90)). 93)), the input / output relationship (formula (99)) is learned using the radial basis function network so that the evaluation (formula (98)) is output. The user number u and the number of intermediate layer neurons, the center of the basis function, the radius of the basis function, and the weight between the basis function and the output layer are stored in the user item evaluation predictor table 40 as the learner parameters.

(3) Prediction Next, a procedure for predicting the evaluation of wine i (formula (101)) unknown to the user u ′ (formula (100)) using this system will be described. Note that the feature amount of wine is determined by any user, and the set of user numbers is represented by equation (102).

Procedure (1)
The U → E feature quantity calculator 24 reads necessary information from the expert item evaluation table 34, and if the equation (83), that is, if the wine i is included in the benchmark wine, the equation (103) is obtained. Skip (2) and go to step (3). If it is Expression (104), the U → E feature quantity calculator 24 reads necessary information from the user item evaluation table 32 and the U → E learner table 36, and the sensor is detected for each user u (Expression (105)). The perceived feature quantity is predicted as Expression (106).

Procedure (2)
The E → U feature quantity calculator 26 uses Expression (107) and Expression (105) calculated in Procedure (1) to obtain Expression (108).

Procedure (3)
The E → U feature quantity calculator 26 reads necessary information from the E → U learner table 38, and uses the feature quantity (formula (109)) calculated up to the procedure (2) to the user u ′ for the wine i. The feature amount perception is predicted as Expression (110).

Procedure (4)
The user item evaluation predicted value calculator 28 reads necessary information from the user item evaluation predictor table 40, and uses the equation (111) calculated in the procedure (3) to evaluate the evaluation of the user u ′ for the wine i by the equation (112). ). The user item evaluation predicted value calculator 28 stores the user number u ′, the wine number i, and the predicted evaluation value (formula (113)) in the user item evaluation predicted value table 42.

(4) Wine recommendation The web server 22 reads necessary information from the item information table 30 and the user item evaluation predicted value table 42 in response to a request from the user terminal 12, and informs the user of wine based on the predicted wine value. To recommend.

B (3) Others This system can be applied to general recommendations for items for which feature quantities can be calculated. In particular, while there are experts or (high cost) sensors that can accurately determine feature quantities, When the user perceives the feature amount, it is useful when the reference is different for each user. For example, this system recommends recommendations for alcoholic beverages such as the above-mentioned wines (beer, sake, shochu, etc.), recommendations for foods such as cheese, recommendations for perfumes, reviews for movies based on reviews, It can be applied to cosmetics recommendations based on reviews.

  As mentioned above, although the form for implementing this invention was demonstrated using the Example, this invention is not limited at all to such an Example, In the range which does not deviate from the summary of this invention, it is with various forms. Of course, it can be implemented.

  The present invention can be used in the manufacturing industry of virtual evaluation systems.

  12 User terminal, 14 Expert terminal, 20 Virtual evaluation system, 22 Web server, 24 User expert feature quantity calculator (U → E feature quantity calculator), 26 Expert user feature quantity calculator (E → U feature quantity calculator), 28 User item evaluation predicted value calculator, 30 item information table, 32 user item evaluation table, 34 expert item evaluation table, 36 user expert learner table (U → E learner table), 38 expert user learner table (E → U learner table), 40 user item evaluation predictor table, 42 user item evaluation prediction value table.

Claims (6)

A virtual evaluation system that specifies and recommends items to be recommended when an item recommendation request is made by a user,
User item information including the feature amount and evaluation value of the item determined by the user for each item for each user, and expert item information including the item feature amount and the evaluation value determined for each item by the expert for each expert And a storage unit for storing information including:
Based on the user item information and the expert item information, for each combination of the user and the expert, a user who outputs the feature amount of the item determined by the expert in response to the input of the item feature amount determined by the user For the expert input / output relationship, the expert user input / output relationship for outputting the feature amount of the item determined by the user in response to the input of the feature amount of the item determined by the expert, and the input of the feature amount of the item by the user A user item evaluation learning unit that learns a user item evaluation input / output relationship that outputs an evaluation value of the item; a learning unit that learns;
A predicted evaluation value prediction unit that predicts a predicted evaluation value for a user item based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship;
Virtual evaluation system with The virtual evaluation system according to claim 1,
The predicted evaluation value prediction unit
Based on the user item information, the expert item information, and the user / expert input / output relationship, for each expert, for each expert for which the feature amount of at least one item has not been determined, the items perceived by the expert A first item feature amount determination unit that predicts feature amounts and determines feature amounts of all items in the expert based on the expert item information and the predicted feature amounts;
Based on the expert / user input / output relationship and the feature amount of all items determined by the first item feature amount determination unit, the feature amount perception of the user's item is predicted and the feature amount perception is used. A second item feature amount determination unit for determining the feature amount of all items in
A prediction unit that predicts the prediction evaluation value based on the user item evaluation input / output relationship and the feature amounts of all items determined by the second item feature amount determination unit;
Having
Virtual evaluation system. The virtual evaluation system according to claim 1,
The predicted evaluation value prediction unit
Based on the user item information, the expert item information, and the user / expert input / output relationship, the feature amount of the item perceived by the expert is predicted for each expert, and the expert item information and the prediction A first item feature amount determination unit that determines feature amounts of all items in the expert based on the feature amount
Based on the expert / user input / output relationship and the feature amount of all items determined by the first item feature amount determination unit, the feature amount perception of the user's item is predicted and the feature amount perception is used. A second item feature amount determination unit for determining the feature amount of all items in
A prediction unit that predicts the prediction evaluation value based on the user item evaluation input / output relationship and the feature amounts of all items determined by the second item feature amount determination unit;
Having
Virtual evaluation system. The virtual evaluation system according to claim 2 or 3,
The first item feature amount determination unit determines feature amounts of all items in the expert by obtaining an average value for each item of the predicted feature amount,
The second item feature amount determination unit determines the feature amount of all items in the user by obtaining an average value for each item of the predicted feature amount.
Virtual evaluation system. A virtual evaluation system according to any one of claims 1 to 4,
A recommendation processing unit for recommending an item based on the predicted evaluation value when an item recommendation request is made from a user;
Virtual evaluation system. A program for causing a computer to function as a virtual evaluation system,
User item information including the feature amount and evaluation value of the item determined by the user for each item for each user, and expert item information including the item feature amount and the evaluation value determined for each item by the expert for each expert And memorize information including
Based on the user item information and the expert item information, for each combination of the user and the expert, a user who outputs the feature amount of the item determined by the expert in response to the input of the item feature amount determined by the user For the expert input / output relationship, the expert user input / output relationship for outputting the feature amount of the item determined by the user in response to the input of the feature amount of the item determined by the expert, and the input of the feature amount of the item by the user Learning a user item evaluation learning unit that learns a user item evaluation input / output relationship that outputs an evaluation value of the item;
Predicting a predicted evaluation value for a user's item based on the user item information, the expert item information, the user / expert input / output relationship, and the expert / user input / output relationship;
A program with

Images