Model generation apparatus, method, and program

Abstract

To reduce evaluations of items having low ratings and reduce the number of evaluations until an optimal solution is obtained, thereby efficiently generating a model and obtaining highly accurate optimal solution.SOLUTION: A model generation apparatus is configured to: obtain an evaluation value based on a subjective view of an evaluator on an item presented; adjust first parameter related to an unsearched portion of a preference distribution of the evaluator and second parameter related to a neighborhood of an optimal value of the preference distribution of a cluster that the evaluator belongs to, in an acquisition function to be used for selecting the item which is to be evaluated, in accordance with reduction characteristics different according to the number of acquisitions of evaluation values; estimate at least a preference distribution of the evaluator based on the evaluation value and the adjusted first and second parameters; train a first model corresponding to the evaluator and a second model corresponding to the cluster, based on a result of the estimation; and select the next item to be evaluated, on the basis of the trained first and second models.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] One aspect of this invention relates to a model generation device, method, and program for generating a model related to subjective data representing a user's personal perception of a target object, for example.

Background Art

[0002] In marketing, it is important to understand what items an individual user prefers and how many people who prefer a certain item exist. If we can understand an individual user's preferences and the group of people with the same preferences, we can understand what products are in demand and what kind of advertisements are effective for what kind of people.

[0003] To understand an individual user's preferences, it is necessary to collect subjective data representing the user's personal perception of an object. However, in some fields, it is difficult to create a mechanism for collecting subjective data. For example, it corresponds to the case of estimating how an individual user feels about something, represented by the evaluation of the five senses, such as estimating an individual user's emotions and degree of favorability from facial expressions, gestures, and voices, or estimating how something like food or drink smells, tastes, and feels delicious.

[0004] If such subjective data representing an individual user's perception can be collected, for example, by converting the facial expression of an avatar or the like according to a person's emotion recognition characteristics, it becomes possible to correctly convey a person's emotions, or to provide a recipe for a dish that the person feels most delicious.

[0005] A known method for collecting subjective data and generating individual user models employs Bayesian optimization. For example, Non-Patent Document 1 describes a mechanism that takes a rating for a certain parameter as input and repeatedly learns how to suggest parameters to explore. In this mechanism, for example, when an individual user rates a dish from a recipe, the next recipe to be rated is suggested, and the individual user rates that recipe as well. This process is repeated until the recipe that maximizes the individual user's rating is output.

[0006] However, the method described in Non-Patent Document 1 has the drawback of requiring a large number of evaluations. Therefore, in order to reduce the number of evaluations, a method has been proposed to construct a subjective model of an individual user by using not only the evaluations of individual users but also the evaluations of multiple other users. For example, Non-Patent Document 2 describes a mechanism that estimates the preference distribution of individual users who are evaluators, and at the same time estimates the preference distribution of the evaluators' cluster, and presents items while taking into account the information of which cluster each individual user belongs to.

[0007] In this system, for example, when an individual user rates a dish from a recipe, clustering is performed by combining the ratings of other users. Then, considering both the individual user's model and the model of the cluster to which the individual user belongs, the next recipe to be rated is output. This process is repeated, with the individual user then rating this recipe, until finally, the recipe with the highest rating from the individual user is output. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P. Adams, and Nando de Freitas, “Taking the Human Out of the Loop: A Review of Bayesian Optimization” Proceedings of the IEEE Vol. 104, No. 1, January 2016 [Non-Patent Document 2] Kotaro Kaihatsu, Masaki Matsubara, Hiroaki Ito, and Atsuyuki Morishima, "Efficient Estimation of Personal and Collective Preference Distributions Based on Gaussian Process Regression," 15th Data Engineering and Information Management Forum (DEIM Forum) 2022, C33-3 [Overview of the Initiative] [Problems that the invention aims to solve]

[0009] However, in the method described in Non-Patent Document 2, individual users end up evaluating low-rated items in order to explore, especially at a stage where the number of evaluations is small. Therefore, when the evaluation target is, for example, aroma, taste, or deliciousness, there is a problem in that individual users have to evaluate data that is unpleasant to them.

[0010] This invention was made in view of the above circumstances, and aims to improve the efficiency of model generation by reducing the evaluation of low-rated items and the number of evaluations required to obtain the optimal solution, thereby providing a technology that can obtain a more accurate optimal solution. [Means for solving the problem]

[0011] To solve the above problems, one embodiment of the model generation apparatus or model generation method according to the present invention acquires subjective evaluation values ​​from an evaluator for a presented item, and adjusts a first parameter relating to the unexplored portion of the evaluator's preference distribution and a second parameter relating to the vicinity of the optimal value of the preference distribution of the cluster to which the evaluator belongs, in an acquisition function used to select the next candidate item for evaluation, according to different reduction characteristics depending on the number of evaluation values ​​acquired. Then, based on the evaluation values ​​and the adjusted first and second parameters, at least the evaluator's preference distribution is estimated, and based on the estimation result, a first model corresponding to the evaluator and a second model corresponding to the cluster are learned, and based on the evaluation values ​​and the learned first and second models, the next candidate item for evaluation is selected.

[0012] According to one aspect of this invention, as the number of acquired evaluation values ​​increases, a first parameter relating to the unexplored portion of the evaluator's preference distribution and a second parameter relating to the vicinity of the optimal value of the preference distribution of the cluster to which the evaluator belongs are adjusted according to different reduction characteristics.

[0013] Therefore, for example, when the number of evaluation values ​​obtained is less than the first value, both the unexplored portion of the evaluator's preference distribution and the area around the optimal value of the preference distribution of the cluster to which the evaluator belongs are given importance when searching for the next candidate item to evaluate. Consequently, the number of items that the evaluator would rate poorly is reduced.

[0014] On the other hand, once the number of evaluation values ​​increases beyond the first value, the area around the optimal value of the cluster's preference distribution is given more weight, and the next evaluation candidate is searched for. This reduces the number of evaluations required to obtain the optimal solution. Note that there are cases where the first and second values ​​are the same, and the area around the optimal value of the cluster's preference distribution is not given more weight.

[0015] Then, when the number of evaluators obtained further increases and exceeds the second value, the vicinity of the optimal value of the preference distribution of the evaluators is emphasized, and the next evaluation candidate item is searched. For this reason, the accuracy of the optimal solution can be improved.

Advantages of the Invention

[0016] That is, according to one aspect of this invention, it is possible to provide a technique for improving the generation efficiency of a model by reducing the evaluation of low-evaluation items and reducing the number of evaluations until an optimal solution is obtained, and obtaining a more accurate optimal solution.

Brief Description of the Drawings

[0017] [Figure 1] FIG. 1 is a diagram showing an example of the configuration of a system including a model generation device according to an embodiment of this invention. [Figure 2] FIG. 2 is a block diagram showing an example of the hardware configuration of a model generation device according to an embodiment of this invention. [Figure 3] FIG. 3 is a block diagram showing an example of the software configuration of a model generation device according to an embodiment of this invention. [Figure 4] FIG. 4 is a flowchart showing an example of the processing procedure and processing content of model generation processing executed by the control unit of the model generation device shown in FIG. 3. [Figure 5] FIG. 5 is a diagram showing an example of evaluation values input by an evaluator for a presented item. [Figure 6] FIG. 6 is a diagram showing an example of the value of a parameter with respect to the number of evaluations set by the model generation device shown in FIG. 3. [Figure 7] FIG. 7 is a diagram showing an example of the change in evaluation values with respect to the number of evaluations in the model generation device shown in FIG. 3, compared with a conventional example.

Modes for Carrying Out the Invention

[0018] Hereinafter, embodiments according to this invention will be described with reference to the drawings.

[0019] [One Embodiment] (Overview) In one embodiment of this invention, in a problem of interactively estimating a preference distribution by a personal user, who is an evaluator, sequentially giving evaluations to presented items from a situation where no evaluation value is given to the items, while estimating the preference distribution of the personal user, the preference distribution of clusters of the evaluator group is estimated, and items are presented in consideration of information on which cluster the personal user belongs to.

[0020] At that time, in an acquisition function for selecting an item to be evaluated next, the weight coefficient given to a term related to exploration of an unexplored part of the preference distribution of the personal user and the weight coefficient given to a term related to exploration around the optimum value of the preference distribution of the cluster to which the personal user belongs are each decreased according to different decreasing characteristics in response to an increase in the number of acquired evaluation values. Note that the optimum value is also included around the optimum value of the above preference distribution.

[0021] Specifically, at a stage where the number of evaluation times is small, by setting both weight coefficients to large values, in addition to a personal model reflecting the preference distribution of the personal user, a cluster model reflecting the preference distribution of the cluster to which the personal user belongs is fully considered, and an item to be evaluated next is selected. On the other hand, at a stage where the number of evaluation times has increased, an item to be presented next is selected mainly based on a personal model reflecting the preference distribution of the personal user.

[0022] By doing so, for example, at a stage where the number of evaluation times is small, in addition to the model of the personal user, the model of the cluster to which the personal user belongs is considered. That is, both the unexplored part of the preference distribution of the evaluator and the vicinity of the optimum value of the preference distribution of the cluster to which the evaluator belongs are emphasized, and the next evaluation candidate items are explored. Therefore, the number of presented evaluation candidate items with low evaluations for the personal user decreases.

[0023] On the other hand, once a sufficient number of evaluations have been conducted, the individual user model is taken into primary consideration; that is, the area around the optimal value of the evaluator's preference distribution is given more weight when selecting the next candidate item for evaluation. As a result, the time it takes for the number of evaluations to decrease and the optimal solution to be obtained is shortened, and the accuracy of the final optimal solution can be improved.

[0024] (Example configuration) (1) System Figure 1 shows an example of the configuration of a system equipped with a model generation device according to one embodiment of the present invention. In Figure 1, SV represents the model generation device, and UT1 to UTn represent user terminals used by individual users who act as evaluators. The model generation device SV is capable of transmitting data between itself and user terminals UT1 to UTn via a network (NW).

[0025] User terminals UT1 to UTn are, for example, smartphones that receive and display items to be evaluated sent from the model generation device SV, and return evaluation values ​​representing subjective feelings entered by the individual user for the displayed items to the model generation device SV. In addition to smartphones, tablet devices, notebook computers, and desktop computers may also be used as user terminals UT1 to UTn.

[0026] A network (NW) comprises, for example, a wide-area network centered on the Internet, and an access network for accessing this wide-area network. Access networks include, but are not limited to, wired or wireless public communication networks, wired or wireless LANs (Local Area Networks), and CATV (Cable Television) networks.

[0027] (2) Model generation device SV Figures 2 and 3 are block diagrams showing examples of the hardware and software configurations of a model generation device SV according to one embodiment of the present invention.

[0028] The model generation device SV is composed of, for example, a server computer or a personal computer. The model generation device SV includes a control unit 1 that uses a hardware processor such as a Central Processing Unit (CPU), and a storage unit having a program storage unit 2 and a data storage unit 3, and a communication interface (hereinafter referred to as I / F) unit 4 are connected to this control unit 1 via a bus 5.

[0029] The communication interface unit 4 has the functions of accessing the network NW and transmitting and receiving data. Under the control of the control unit 1, it transmits the items to be evaluated to user terminals UT1 to UTn, and receives evaluation values ​​entered by individual users for the items from user terminals UT1 to UTn.

[0030] The program storage unit 2 is configured, for example, as a storage medium, by combining a non-volatile memory that can be written to and read at any time, such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), with a non-volatile memory such as ROM (Read Only Memory). In addition to middleware such as an OS (Operating System), it stores various programs necessary to execute various control processes according to one embodiment of this invention.

[0031] The data storage unit 3 is configured, for example, as a storage medium, by combining a non-volatile memory that can be written to and read at any time, such as an HDD or SSD, with a volatile memory such as RAM (Random Access Memory). It includes an evaluation value storage unit 31 and a model storage unit 32 as storage areas necessary for carrying out one embodiment of this invention.

[0032] The evaluation value storage unit 31 stores the evaluation values ​​received from user terminals UT1 to UTn, associating them with the identification information of the individual user who sent the data (user ID) or the identification information of user terminals UT1 to UTn (terminal ID).

[0033] The model memory unit 32 consists of an individual model representing the preference distribution for each individual user and multiple cluster models. The cluster models are generated for each of the multiple clusters formed by clustering the preference distributions of multiple individual users that constitute the evaluator group.

[0034] The control unit 1 includes, as a processing function according to one embodiment of the present invention, an evaluation value acquisition processing unit 11, a parameter adjustment processing unit 12, a model learning processing unit 13, an evaluation candidate selection processing unit 14, and an evaluation candidate output processing unit 15.

[0035] Each of these processing units 11 to 15 is implemented by having the hardware processor of the control unit 1 execute the application program stored in the program storage unit 2. Note that some or all of the above processing units 11 to 15 may be implemented using hardware such as LSI (Large Scale Integration) or ASIC (Application Specific Integrated Circuit).

[0036] The evaluation value acquisition processing unit 11 receives evaluation values ​​entered by individual users for presented items, transmitted from user terminals UT1 to UTn, via the communication interface unit 4. The received evaluation values ​​are then stored in the evaluation value storage unit 31, associated with the user ID or terminal ID representing the source of transmission.

[0037] The parameter adjustment processing unit 12 adjusts the parameters used to adjust the balance between searching for and utilizing items, which are included in the acquisition function used to select the next candidate item to be presented for evaluation, for each individual user being evaluated. These parameters are reduced according to different reduction characteristics as the number of evaluation values ​​stored in the evaluation value storage unit 31 increases, i.e., the number of evaluation values ​​acquired (also called the number of evaluations). An example of this parameter adjustment process will be explained in the operation example.

[0038] The model learning processing unit 13 estimates the preference distribution of each individual user and the preference distribution of the cluster to which the individual data belongs, using the latest evaluation value stored in the evaluation value storage unit 31 and the adjusted parameters, respectively, for each individual user to be evaluated. Based on these estimation results, it learns the corresponding individual model and the cluster model to which the individual user belongs.

[0039] The evaluation candidate selection processing unit 14, for each individual user being evaluated, uses the latest evaluation value stored in the evaluation value storage unit 31 and the trained individual model and cluster model stored in the model storage unit 32 to select the next item to present, taking into consideration the balance between a search process to select an unknown portion of the individual user's preference distribution and an utilization process to select items that the individual user likes. An example of this item selection process will also be explained in the operation example.

[0040] The evaluation candidate output processing unit 15 transmits the above-mentioned items selected as the next evaluation target from the communication I / F unit 4 to the user terminal UT used by the individual user being evaluated.

[0041] (Example of operation) Next, we will explain an example of the operation of the model generation device SV configured as described above. Figure 4 is a flowchart showing an example of the processing procedure and processing content of the model generation process executed by the control unit 1 of the model generation device SV.

[0042] (1) Obtaining evaluation values ​​for the presented items In step S10, when the control unit 1 of the model generation device SV detects a model generation instruction sent from, for example, user terminals UT1 to UTn or a management terminal (not shown) used by the system administrator, it first selects one individual user to be evaluated from among a group of individual users pre-registered as an evaluator group in step S11. Then, under the control of the evaluation candidate output processing unit 15, in step S12, the control unit 1 of the model generation device SV sends, for example, a pre-prepared item from the communication I / F unit 4 to the user terminal (for example, UTk) corresponding to the selected individual user.

[0043] When the above item is sent from the model generation device SV to the user terminal UTk, the item is received and displayed on the display device. In this state, when the individual user, who is the evaluator, enters their evaluation value for the above item on the input device, the entered evaluation value is sent from the user terminal UTk to the model generation device SV.

[0044] Examples of the items mentioned above include cooking recipes. Figure 5 shows an example of a recipe presented and an example of the evaluation value that an individual user provided for this recipe. In this example, the evaluation value is represented by a value selected from, for example, 10 levels, which indicate the degree of preference.

[0045] In addition, other items, such as the color tone of an image, may also be applied. In this case, the model generation device SV selects an image of any color from several pre-prepared sample images of different colors and sends it to the user terminal UTk. The evaluator then selects and inputs the evaluation value that best matches their own perception from several evaluation values ​​representing the degree of color preference, and the selected evaluation value is sent back from the user terminal UTk to the model generation device SV.

[0046] After the transmission of the above items is complete, when the above evaluation value is returned from the user terminal UTk, the control unit 1 of the model generation device SV receives the above evaluation value via the communication I / F unit 4 in step S13 under the control of the evaluation value acquisition processing unit 11, and stores it in the evaluation value storage unit 31 in association with the user ID or terminal ID of the individual user who is the evaluator.

[0047] (2) Parameter adjustment When the above evaluation value is received, the control unit 1 of the model generation device SV, under the control of the parameter adjustment processing unit 12, performs the following parameter adjustment process in step S14.

[0048] In other words, the item selection process uses an acquisition function consisting of terms representing the expected value and variance of the evaluator's preference distribution, and a term representing the expected value of the preference distribution of the cluster to which the evaluator belongs. The evaluator's preference distribution and the cluster's preference distribution can be represented, for example, by a Gaussian process regression model. The equation representing the acquisition function is shown below.

number

[0049] However, x is the item's feature, y is the evaluation value, D is the set of all evaluation values, E is the mean, V is the variance, ω is the evaluator, C k is the kth cluster, nc is the number of clusters, γ ωk This value represents a value of 1 in the cluster to which a user belongs and 0 in the cluster to which they do not belong.

[0050] Furthermore, α1 and α2 are weight coefficients used as the first and second parameters, respectively. Of these, α1 is reflected in the term that explores the unexplored portion of the evaluator's preference distribution, and α2 is reflected in the term that explores the area around the optimal value of the cluster's preference distribution.

[0051] The parameter adjustment processing unit 12 first determines the number of evaluation values ​​acquired from the start of the evaluation to the present time, i.e., the number of evaluation value acquisitions (or evaluation count) t, based on the accumulated state of evaluation values ​​stored in the evaluation value storage unit 31. Then, the parameter adjustment processing unit 12 adjusts the weight coefficients α1 and α2 so that they decrease according to a predetermined decrease characteristic as the number of evaluation value acquisitions t increases. Note that the weight coefficient of the term in the acquisition function that searches around the optimal value of the evaluator's preference distribution is always set to "1".

[0052] Below is an example of the formula used for this setting.

number

[0053] Figure 6 shows an example of the decrease in weight coefficients α1 and α2 with respect to an increase in the number of times evaluation values ​​are obtained, when weight coefficients α1 and α2 are set according to the above formula.

[0054] Furthermore, a monotonically decreasing function may be used to set the weight coefficients α1 and α2, or other functions with decreasing characteristics may be used. In one embodiment, the case in which the weight coefficients α1 and α2 are set each time a new evaluation value is obtained is described, but the weight coefficients α1 and α2 may also be set each time a predetermined number of evaluation values ​​are obtained.

[0055] (3) Model training The control unit 1 of the model generation device SV then performs the model learning process under the control of the model learning processing unit 13 as follows.

[0056] Specifically, in step S15, the model learning processing unit 13 first estimates the preference distribution of the individual user being evaluated based on the latest evaluation value stored in the evaluation value storage unit 31 and the weight coefficient α1 set by the parameter adjustment unit. This estimation of the individual user's preference distribution is performed by recalculating the expected value and variance of the evaluation of the presented item by the individual user who is the evaluator. Subsequently, in step S16, the model learning processing unit 13 learns and updates the corresponding individual model stored in the model storage unit 32 based on the estimation result of the individual user's preference distribution.

[0057] Next, in step S17, the model learning processing unit 13 identifies clusters that have preference distributions highly similar to the preference distribution of the individual user. In other words, it identifies the cluster to which the individual user belongs. Then, in step S18, the model learning processing unit 13 estimates the preference distribution for the cluster to which the individual user belongs. This estimation of the preference distribution for a cluster is performed by recalculating the expected value and variance of the evaluation of the presented items for that cluster. Subsequently, in step S19, the model learning processing unit 13 learns and updates the corresponding cluster model stored in the model storage unit 32 based on the estimation result of the preference distribution for the cluster.

[0058] (4) Selection and presentation of items to be evaluated Once the model update process described above is complete, the control unit 1 of the model generation device SV, under the control of the evaluation candidate selection processing unit 14, performs the following process in step S20 to select the evaluation candidate items to be presented to the individual user next.

[0059] In other words, the evaluation candidate selection processing unit 14 uses in combination the acquisition function consisting of terms representing the expected value and variance of the evaluator's preference distribution, as shown in equation (1) above, and the acquisition function consisting of a term representing the expected value of the preference distribution of the cluster to which the evaluator belongs. This allows the unit to select an item with features that maximizes the acquisition function while considering the balance between searching for and utilizing evaluation candidates, and the selected item is designated as the next evaluation candidate item to be presented.

[0060] In one embodiment, the above selection process is achieved by using a combination of an individual model that reflects the preference distribution of individual users estimated by the model learning processing unit 13, i.e., a term of the acquisition function representing the preference distribution of the evaluator, and a cluster model that reflects the preference distribution of the cluster to which the individual user belongs, i.e., a term of the acquisition function representing the preference distribution of the cluster to which the evaluator belongs, and applying the latest evaluation values ​​obtained by the evaluation value acquisition processing unit 11 to these models.

[0061] Here, the individual model and cluster model described above are trained based on the estimated preferences of individual users and clusters, respectively, and each estimation result reflects the weight coefficients α1 and α2, which are used to adjust the balance between exploring and utilizing evaluation candidates. Therefore, the degree to which the individual model and cluster model contribute to the selection of the next item changes depending on the number of times evaluation values ​​are obtained.

[0062] For example, when the number of times an evaluation value has been obtained is less than the first value determined by the decreasing characteristics of the weight coefficients α1 and α2, the unexplored portion of the individual user's preference distribution and the area around the optimal value of the preference distribution of the cluster to which the individual user belongs are both given importance when searching for the next candidate item for evaluation. Therefore, candidate items for evaluation are selected by considering both the unexplored portion of the individual user's preference distribution and the area around the optimal value of the preference distribution of the cluster to which the user belongs. Consequently, the presentation of items that would receive a low evaluation from the individual user is kept to a minimum.

[0063] Figure 7 shows an example of a simulation result of the change in the user's evaluation value with respect to the number of times evaluation values ​​are obtained (number of evaluations). A shows the change according to this embodiment, and B shows the change when the weight coefficients α1 and α2 are fixed to constant values. As shown in this example, this embodiment makes it possible to reduce the presentation of low-rated items at the stage when the number of evaluations is small.

[0064] On the other hand, as the number of evaluations increases beyond the first value mentioned above, emphasis is placed on the area around the optimal value of the cluster's preference distribution, and the next candidate item for evaluation is searched for. As a result, the number of evaluations required to obtain the optimal solution decreases.

[0065] Furthermore, as the number of evaluations increases, and exceeds a second value determined by the decreasing characteristics of the weight coefficients α1 and α2, emphasis is placed on the area around the optimal value of the individual user's preference distribution, and the next candidate items for evaluation are searched for. Therefore, it becomes possible to improve the accuracy of the optimal solution for the presented items.

[0066] The control unit 1 of the model generation device SV transmits the selected evaluation candidate items as described above to the user terminal UTk used by the individual user via the communication I / F unit 4 in step S12, under the control of the evaluation candidate output processing unit 15.

[0067] (5) Repeat the evaluation process The control unit 1 of the model generation device SV repeatedly executes the series of processes described above in steps S12 to S20. Then, in step S21, it determines whether the evaluation process for one individual user selected in step S11 has been completed. This determination is made, for example, by determining whether the number of evaluations has reached a predetermined value, or whether the individual user's evaluation value for the presented items has reached a value equal to or greater than a predetermined value corresponding to the optimal solution.

[0068] If the evaluation process for the individual user is completed based on the above determination, the control unit 1 of the model generation device SV proceeds to step S22. In step S22, the control unit 1 of the model generation device SV determines whether the evaluation process for all pre-registered evaluators has been completed.

[0069] If, as a result of this determination, there are still evaluators who have not yet performed an evaluation, the control unit 1 of the model generation device SV returns to step S11 to select the next individual user, and then performs the evaluation process for the selected individual user in steps S12 to S22.

[0070] Thereafter, the control unit 1 of the model generation device SV will similarly perform the evaluation process for each of the evaluators being evaluated, and will terminate the series of model generation processes once the evaluation process for all evaluators is complete.

[0071] (effect) As described above, in one embodiment, an acquisition function is used that has a term representing the preference distribution of an individual user who is an evaluator and a term representing the preference distribution of the cluster to which the individual user belongs. When selecting the next item to present to the evaluator, the weight coefficient α1 given to the term that explores the unexplored portion of the individual user's preference distribution and the weight coefficient α2 given to the term that explores the area around the optimal value of the cluster's preference distribution are adjusted to decrease according to a decrease characteristic set so that the decrease in weight coefficient α2 becomes larger than that of weight coefficient α1 as the number of times evaluation values ​​are acquired increases.

[0072] Therefore, for example, in the early stages when the number of evaluation values ​​obtained is small, both the unexplored portion of the individual user's preference distribution and the area around the optimal value of the preference distribution of the cluster to which that individual user belongs are given importance when searching for the next candidate item to evaluate. As a result, items that reflect the area around the optimal value of the cluster model's preference distribution are proactively selected, and it is possible to minimize the presentation of items that would receive low ratings due to the unexplored portion for the individual user.

[0073] On the other hand, as the number of evaluations increases, emphasis is placed on the area around the optimal value of the cluster preference distribution, and the next evaluation candidate is searched for. Therefore, the number of evaluations required to obtain the optimal solution can be reduced, thereby reducing the processing load and time required to generate the model.

[0074] Once a sufficient number of evaluations have been conducted, the system prioritizes the area around the optimal value of the individual user's preference distribution when searching for the next candidate item to be evaluated. Therefore, the next candidate item is selected from near the optimal value of the individual user's preference distribution. Consequently, it becomes possible to improve the accuracy of the optimal solution for the presented items.

[0075] [Other embodiments] (1) In one embodiment, the clustering of individual users and the calculation of the cluster preference distribution are performed each time a new evaluation value is obtained. However, the recalculation of the clustering of individual users and the cluster preference distribution may be performed each time a predetermined evaluation value is obtained. The same applies to the setting process of the weight coefficients α1 and α2.

[0076] (2) In one embodiment, the example described was that all the processes provided by the model generation device SV are executed on a server computer located on the Web or in the cloud. However, the above processes may also be executed on a personal computer used by the system administrator in addition to the server computer. Furthermore, each process provided by the model generation device SV may be distributed across multiple server computers or personal computers that can transfer data to each other via a network NW.

[0077] (3) In addition, the procedures and content of the processing performed by each processing unit of the model generation device, the characteristics of parameter changes, the types of items to be evaluated, etc., can be modified in various ways without departing from the spirit of this invention.

[0078] Although embodiments of this invention have been described in detail above, the above description is merely illustrative in all respects. It goes without saying that various improvements and modifications can be made without departing from the scope of this invention. In other words, when implementing this invention, specific configurations may be adopted as appropriate depending on the embodiment.

[0079] In short, this invention is not limited to the embodiments described above, and in the implementation stage, the components can be modified and materialized without departing from the gist of the invention. Furthermore, various inventions can be formed by appropriately combining the multiple components disclosed in the embodiments. For example, some components may be deleted from all the components shown in the embodiments. Moreover, components from different embodiments may be appropriately combined. [Explanation of symbols]

[0080] SV...Model Generator UT1~UTn,UTk…User terminals NW...Network 1…Control Unit 2…Program memory 3…Data storage unit 4…Communication I / F section 5... Bus 11…Evaluation value acquisition processing unit 12…Parameter adjustment processing unit 13…Model Learning Processing Unit 14…Evaluation Candidate Selection Processing Unit 15…Evaluation candidate output processing unit 31...Evaluation value storage unit 32...Model memory unit

Claims

1. A first processing unit that obtains subjective evaluation values ​​from evaluators for the presented items, A second processing unit adjusts, among the acquisition function used to select the next candidate item for evaluation, a first parameter relating to the unexplored portion of the evaluator's preference distribution and a second parameter relating to the vicinity of the optimal value of the preference distribution of the cluster to which the evaluator belongs, according to different reduction characteristics depending on the number of evaluation values ​​obtained. A third processing unit that estimates at least the preference distribution of the evaluators based on the evaluation value and the adjusted first and second parameters, and learns a first model corresponding to the evaluators and a second model corresponding to the clusters based on the estimation results, A fourth processing unit that selects the next candidate item for evaluation based on the evaluation value and the learned first model and second model. A model generation device equipped with the following features.

2. The model generation apparatus according to claim 1, wherein the second processing unit reduces the first parameter and the second parameter according to the reduction characteristic set such that the reduction amount of the first parameter becomes larger than that of the second parameter in response to an increase in the number of evaluation values ​​obtained.

3. The fourth processing unit searches for the next evaluation candidate based on the evaluation value and the learned first and second models, when the number of obtained evaluation values ​​is less than the first value, by giving importance to both the unexplored portion of the evaluator's preference distribution and the area around the optimal value of the preference distribution of the cluster to which the evaluator belongs; when the number of obtained evaluation values ​​exceeds the first value, by giving importance to the area around the optimal value of the cluster's preference distribution; and when the number of obtained evaluators further increases and exceeds the second value, by giving importance to the area around the optimal value of the evaluator's preference distribution. The model generation apparatus according to claim 1.

4. A model generation method executed by an information processing device, The process of obtaining subjective evaluation values ​​from evaluators for the presented items, The acquisition function used to select the next candidate item for evaluation includes a process of adjusting a first parameter relating to the unexplored portion of the evaluator's preference distribution and a second parameter relating to the vicinity of the optimal value of the preference distribution of the cluster to which the evaluator belongs, according to different reduction characteristics depending on the number of evaluation values ​​obtained. A process of estimating at least the preference distribution of the evaluators based on the evaluation values ​​and the adjusted first and second parameters, and learning a first model corresponding to the evaluators and a second model corresponding to the clusters based on the estimation results, A process of selecting the next candidate item for evaluation based on the evaluation value and the learned first and second models. A model generation method comprising the following features.

5. A program that causes a processor in a model generation device to execute at least one of the processes performed by the first to fourth processing units in a model generation device according to any one of claims 1 to 3.

Images