JP5503509B2 - Trust relationship strength estimation device, trust relationship strength estimation method, evaluation value prediction device, evaluation value prediction method, and program - Google Patents

Description

  The present invention relates to a trust relationship strength estimation device, a trust relationship strength estimation method, an evaluation value prediction device, an evaluation value prediction method, and a program, and more particularly to a trust relationship strength estimation device that estimates the strength of trust relationships between users, The present invention relates to a relationship strength estimation method, a program, and an evaluation value prediction apparatus, an evaluation value prediction method, and a program for predicting an evaluation value by a user for an item.

  In recent years, product review sites have attracted attention. The product review site allows users to write reviews and give ratings to various products. Some product review sites allow users to create social trust relationships with other users. The trust relationship created by the user indicates that user A recognizes that user B's reviews and ratings can be trusted.

  Moreover, a product suitable for a certain user can be recommended by the collaborative filtering algorithm. Recently, several algorithms have been proposed that use trust relationships to predict user preferences. For example, an algorithm for making a recommendation by combining a trust relationship and collaborative filtering has been proposed (for example, Non-Patent Document 1). Further, an algorithm for predicting user evaluation by incorporating a trust relationship into an existing matrix decomposition method has been proposed (Non-Patent Document 2).

Jamali, M. and Ester, M., “TrustWalker: a random walk model for combining trust-based and item-based recommendation”, KDD '09, pp.397-406, 2009. Ma, H., Lyu, M.R. and King, I., “Learning to recommend with trust and distrust relationships”, RecSys '09, pp.189-196, 2009.

  In the techniques described in Non-Patent Documents 1 and 2, when there is a trust relationship, it is assumed that the strength of the trust relationship is the same among all users. However, in reality, even if there is a trust relationship between the two users, there are some users who have a strong trust relationship and some users who are weak. , Do not always have the same opinion. Therefore, there is a problem that when a product is recommended assuming that the strength of the trust relationship between users is the same, a suitable recommendation cannot be performed.

  The present invention has been made in order to solve the above-described problems, and is a trust relationship strength estimation device, a trust relationship strength estimation method, an evaluation value prediction device, and an evaluation value prediction that can estimate the strength of a trust relationship between users. It is an object to provide a method and a program.

  In order to achieve the above object, a trust relationship strength estimation apparatus according to a first aspect of the present invention relates to a plurality of users and a plurality of items, user information indicating the user, item information indicating the item, and Storage means for storing a set of observation data including evaluation values evaluated for the item; a user feature parameter representing each feature of the plurality of users; an item feature parameter representing each feature of the plurality of items; Initial value setting means for setting an initial value of each of the strength parameters representing the strength of the trust relationship between the users, the user characteristic parameter of the user information of the observation data for each of the observation data, the observation data The item feature parameter of the item information, the user indicated by the user information and other users The error between the evaluation value of the observation data and the evaluation value calculated based on the evaluation value of each of the other users with respect to the item information obtained from the set of the observation data and the intensity parameter between Estimating means for updating each value of the user feature parameter of each of the plurality of users, the item feature parameter of each of the plurality of items, and each of the intensity parameters between the users by an evaluation function; The user feature parameter, the item feature parameter, and the intensity parameter are repeatedly updated for the set of observation data by repeating the updating of the user feature parameter, the item feature parameter, and the intensity parameter by the estimation unit. When optimized, the user feature parameter, the item feature parameter Is constructed over data, and an output means for outputting the intensity parameter, it includes.

  The reliability relationship strength estimation method according to the second invention is the user information indicating the user, the item information indicating the item, and the evaluation evaluated by the user for the item for a plurality of users and a plurality of items. A trust relationship strength estimation method in a trust relationship strength estimation apparatus including a storage means that stores a set of observation data including values, an initial value setting means, an estimation means, and an output means, the trust relationship strength estimation apparatus Is a user feature parameter that represents each feature of the plurality of users, an item feature parameter that represents each feature of the plurality of items, and an intensity parameter that represents the strength of the trust relationship between the users. Setting an initial value of each of the observation data, and by the estimating means, the observation data for each of the observation data. Obtained from the user feature parameter of the user information, the item feature parameter of the item information of the observation data, the intensity parameter between the user indicated by the user information and another user, and the set of the observation data The user of each of the plurality of users by an evaluation function using an error between an evaluation value calculated based on each evaluation value of the other user information for the item information and an evaluation value of the observation data Updating a value of each of the feature parameter, the item feature parameter of each of the plurality of items, and the intensity parameter of each of the users, and the user feature parameter by the estimating means by the output means , Repeatedly updating the item feature parameter and the intensity parameter. And outputting the user feature parameter, the item feature parameter, and the intensity parameter when the user feature parameter, the item feature parameter, and the intensity parameter are optimized for the set of observation data. And executing.

  According to the trust relationship strength estimation apparatus according to the first invention and the trust relationship strength estimation method according to the second invention, the initial value setting means allows the user feature parameters representing the features of the plurality of users, An initial value is set for each of the item feature parameter representing the feature of each item and the strength parameter representing the strength of the trust relationship between the users.

  Then, for each of the observation data, the user feature parameter of the user information of the observation data, the item feature parameter of the item information of the observation data, the user indicated by the user information and other users by the estimation means An error between an evaluation value calculated based on each evaluation value of each of the other users with respect to the item information obtained from the set of observation data and the intensity parameter between and the observation data The value of each of the user feature parameter of each of the plurality of users, the item feature parameter of each of the plurality of items, and each of the intensity parameters between the users is updated according to the evaluation function used.

  By repeating the updating of the user feature parameter, the item feature parameter, and the intensity parameter by the estimation unit by the output unit, the user feature parameter, the item feature parameter, and When the intensity parameter is optimized, the user characteristic parameter, the item characteristic parameter, and the intensity parameter are output.

  As described above, the evaluation function using the error between the evaluation value calculated based on the user characteristic parameter, the item characteristic parameter, the intensity parameter, and the evaluation value of other users and the evaluation value of the observation data is used. By optimizing the feature parameters, item feature parameters, and strength parameters, the strength of the trust relationship between users can be estimated.

  According to the first invention, the calculated evaluation value of the user information u with respect to the item information i is related to the degree of relevance between the user information u and the potential feature k with respect to a plurality of potential features k. , The total value obtained by adding the products obtained by multiplying the potential feature k and the relevance of the item information i, and the user information v of the user trusted by the user indicated by the user information i obtained in advance. For each, a total value obtained by adding the products of the strength parameter between the user information i and the user information v and the evaluation value of the user information v for the item information i is the value of the user information v. The quotient obtained by dividing the intensity parameter between the user information i and the user information v by the total value added to each of them can be calculated by weighted addition.

  Further, the set of observation data includes a plurality of observation data including an average value of evaluation values for the user information, the item information, and the item information generated for each of the item information, and indicating a virtual user, Each of the plurality of users may trust the virtual user. Thereby, even if there is a user who does not trust any user, the strength of the trust relationship between the users can be estimated.

  An evaluation value predicting apparatus according to a third aspect of the invention relates to the reliability relation strength estimation apparatus according to the first aspect of the invention, and the user feature parameter, the item feature parameter, and the strength parameter output by the trust relation strength device. The user of the user information stored in the parameter storage means stores the parameter storage means for storing and the evaluation value of the prediction target user of the plurality of users for the prediction target item of the plurality of items. And an item evaluation prediction means for predicting using the feature parameter, the item feature parameter, and the intensity parameter.

  An evaluation value prediction method according to a fourth invention includes a reliability relationship strength estimation device according to the first invention, and the user feature parameter, the item feature parameter, and the strength parameter output by the trust relationship strength device. An evaluation value prediction method in an evaluation value prediction apparatus including a parameter storage means for storing and an item evaluation prediction means, wherein the evaluation value prediction apparatus uses the initial value setting means to characterize each of the plurality of users. Setting an initial value of each of a user feature parameter to be represented, an item feature parameter representing a feature of each of the plurality of items, and an intensity parameter representing the strength of a trust relationship between the users; and For each of the data, for the user information and the item information of the observation data, The user feature parameter of the user information, the item feature parameter of the item information, the intensity parameter between the user indicated by the user information and another user, and the item information obtained from the set of observation data The user characteristic parameters of each of the plurality of users, the plurality of user parameters, and an evaluation function using an error between the evaluation value calculated based on each evaluation value of other user information and the evaluation value of the observation data Updating the value of each of the item feature parameters of each of the items and each of the intensity parameters between the users, and by the output means, the user feature parameters by the estimation means, the item feature parameters, And repeatedly updating the intensity parameter, Output the user feature parameter, the item feature parameter, and the strength parameter when the user feature parameter, the item feature parameter, and the strength parameter are optimized, and the item evaluation prediction unit The user characteristic parameter of the user information stored in the parameter storage means, the evaluation value of the prediction target user of the plurality of users with respect to the prediction target item of the plurality of items, the item And performing prediction using the characteristic parameter and the intensity parameter.

  According to the evaluation value prediction apparatus according to the third invention and the evaluation value prediction method according to the fourth invention, the initial value setting means, the user characteristic parameter representing each characteristic of the plurality of users, the plurality of items The initial value of each of the item feature parameters representing the respective features and the strength parameters representing the strength of the trust relationship between the users is set.

  Then, for each of the observation data, the user feature parameter of the user information of the observation data, the item feature parameter of the item information of the observation data, the user indicated by the user information and other users by the estimation means Between the evaluation value calculated based on the evaluation value of each of the other user information with respect to the item information obtained from the set of observation data and the intensity parameter between and the evaluation value of the observation data Updating each value of the user feature parameter of each of the plurality of users, the item feature parameter of each of the plurality of items, and each of the intensity parameters between the users by an evaluation function using .

  By repeating the updating of the user feature parameter, the item feature parameter, and the intensity parameter by the estimation unit by the output unit, the user feature parameter, the item feature parameter, and When the intensity parameter is optimized, the user characteristic parameter, the item characteristic parameter, and the intensity parameter are output.

  The evaluation value of the prediction target user of the plurality of users with respect to the prediction target item of the plurality of items by the item evaluation prediction unit is stored in the user information stored in the parameter storage unit. Prediction is performed using the user feature parameter, the item feature parameter, and the intensity parameter.

  As described above, the evaluation function using the error between the evaluation value calculated based on the user characteristic parameter, the item characteristic parameter, the intensity parameter, and the evaluation value of other users and the evaluation value of the observation data is used. By optimizing the feature parameters, item feature parameters, and strength parameters, the strength of the trust relationship between users can be estimated. Moreover, the user's evaluation value for the prediction target item can be accurately predicted using the user feature parameter, the item feature parameter, and the intensity parameter.

  The item evaluation predicting means according to a third invention is configured such that an evaluation value of the user information u with respect to the item information i is obtained from the user information u and the potential feature k with respect to a plurality of potential features k. The total value obtained by adding the product obtained by multiplying the degree of relevance by the degree of relevance of the potential feature k and the item information i, and the user trusted by the user indicated by the user information i obtained in advance For each information v, a total value obtained by adding the product of the strength parameter between the user information i and the user information v and the evaluation value of the user information v for the item information i is the user For each information v, the quotient obtained by dividing the user information i and the intensity parameter between the user information v by the total value added together may be calculated by weighted addition. Kill.

  A program according to a fifth aspect of the invention is a program for causing a computer to function as each means of the above-described reliability relationship strength estimating apparatus or each means of the above-described evaluation value predicting apparatus.

  As described above, according to the reliability relationship strength estimation device, the reliability relationship strength estimation method, the evaluation value prediction device, the evaluation value prediction method, and the program of the present invention, the user feature parameter, item feature parameter, strength parameter, and others By optimizing user feature parameters, item feature parameters, and intensity parameters using an evaluation function using an error between an evaluation value calculated based on the user's evaluation value and an evaluation value of observation data, The effect that the strength of the trust relationship between users can be estimated is obtained.

It is the schematic which shows the structure of the trust relationship strength estimation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the trust relationship strength estimation process routine in the trust relationship strength estimation apparatus which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the evaluation value prediction apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the content of the goods evaluation value prediction process routine in the evaluation value prediction apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows the result of having estimated the evaluation value of goods.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Overview>
First, the principle of the present invention will be described. Assume that the product evaluation data as a set of observation data is S = (U, I, R). U represents a user set, and I represents a product set. R represents a set of user evaluation values for the product. Each evaluation value is represented by an integer from 1 to 5, but may be represented by an arbitrary real number. The product evaluation data S can be represented by an M × N matrix. M is the number of users and N is the number of products. An element r _{ui of} this matrix is an evaluation value given to the product i by the user u. An element of product evaluation data can be represented by (u, i, r _ui ).

  Here, the implementation of the proposed method will be described based on matrix decomposition, but this proposed method can be used in combination with other collaborative filtering techniques such as cosine similarity instead of matrix decomposition.

Matrix decomposition is a technique for decomposing a matrix into the product of two matrices. When this method is used as an evaluation value prediction method, a matrix R having evaluation r _ui as elements (where u is a user index (1 ≦ u ≦ M) and i is a product index (1 ≦ i ≦ N). ), And r _ui represents the (u, i) element of R) is decomposed into the product of two matrices P and Q. As a result of matrix decomposition, user characteristics and product characteristics can be understood. The user index is an example of user information, and the product index is an example of item information.

In order to estimate the strength of the trust relationship between users, the above matrix decomposition method is extended as described below. First, the trust relationship between users is represented by an M × M matrix called G. When the user u trusts the user v (1 ≦ u, v ≦ M), g _uv = 1, and when there is no trust relationship, g _uv = 0. Here, g _uv represents the (u, v) element of G. The strength of the trust relationship between users is represented by an M × M matrix called S, and the elements of S are represented by s _uv .

  It is assumed that there are two factors in the evaluation given to the product by the user. The first factor is user preferences (features) and product features. User preferences and product features can be calculated with matrix decomposition and other collaborative filtering techniques. The second factor is an influence from a trusted user. When the user u trusts the user v, it is highly possible that the evaluations of the two people are similar to the same product. However, the proposed method of the present invention assumes that a user is not equally affected by all the trusted users.

  Based on the above idea, a standard matrix decomposition method is extended to propose a model represented by the following equation (1).

  Where α is a parameter that controls the contribution of two factors. Here, α is set to an arbitrary value in advance, but an optimum value of α may be estimated and set using machine learning technology.

The first term of the equation (1) is a term representing the first factor, and the evaluation value is estimated using the potential feature. p _uk is a parameter representing the strength of association between the user u and the potential feature k (where k is a feature index and 1 ≦ k ≦ K), and is an example of a user feature parameter. q _ki is a parameter representing the strength of the relationship between the potential feature k and the product i, and is an example of an item feature parameter. Further, if P is an M × K matrix having p _uk as an (u, k) element and Q is a K × N matrix having q _ki as an (k, i) element, then R = PQ holds. Although potential features are estimated using matrix decomposition, as described above, this first term may be replaced by cosine similarity or other collaborative filtering techniques. For example, when replaced with cosine similarity, the first term predicts the user's evaluation of the new product using the similarity between the product in the user's evaluation history and the new product. However, even if the first term is replaced with cosine similarity or another collaborative filtering technique, the second term does not change and the strength of the trust relationship can be estimated.

The second term of the above equation (1) is a term representing the second factor, and calculates the influence of the evaluation value of other trusted users. Here, not only g _uv but also s _uv is used. The reason for this is to consider that the trust relationship is not a binary value of 1 (there is a trust relationship) or 0 (no trust relationship), but has different strengths of trust relationship for each combination of users. Moreover, it is for estimating the evaluation value by estimating the strength of the trust relationship. s _uv is learned by machine learning techniques in the same way as p _uk and q _ki .

Thus, the above equation (1) indicates that the evaluation r _ui of the user u with respect to the product i has the relationship P _uk between the user u and the potential feature k, and the potential feature k and the product i. A model that assumes that a value obtained by multiplying the value of the related strength q _ki with respect to all potential features k is a value corrected in consideration of the strength s _uv of the trust relationship between users. It is.

If a user does not trust any user, the numerical value in the second term is zero, and there is a high possibility that the evaluation prediction will be missed. In order to solve this problem, the virtual user z is added to the user set, and the evaluation (z, i, r _zi ) for each product i (1 ≦ i ≦ N) of the virtual user z is used as the product evaluation data S. Add to Assume that all users trust virtual user z. Further, it is assumed that the virtual user z has given the evaluation value r _zi to all the products i, and the evaluation value r _zi is an average value of the evaluation values given to the product i. Thereby, even when there is no reliable user, the evaluation value can be appropriately predicted.

Further, when estimating the parameters s _uv , p _uk , and q _ki , the parameter s is set so that the evaluation value r _ui estimated according to the above equation (1) is close to the evaluation value r _ui obtained from the evaluation data of the product. _uv, p _uk, to estimate the q _ki. For example, the parameters s _uv , p _uk , and q _ki can be estimated by an optimization problem using an evaluation function expressed by the following equation (2).

However, O is a set of product evaluation data, that is, observation data (u, i, r _ui ).

For example, the parameters s _uv , p _uk , and q _ki can be estimated using a gradient descent method. The parameters s _uv , p _uk , and q _ki are initialized with random values, and the parameters s _uv , p _uk , and q _ki are optimized using an update rule expressed by the following equation (3).

  here,

Is an error of the estimated evaluation value, and γ is a meta parameter of learning. ← means that the value on the right side is assigned to the parameter on the left side.

The strength s _uv of the trust relationship between users can be obtained based on the result of optimization of the parameters s _uv , p _uk , q _ki using the above evaluation function.

<System configuration>
Next, taking as an example a case where the present invention is applied to a trust relationship strength estimation device that estimates the strength of trust relationships between users by using evaluation data of products, which is a set of observation data, of the first invention. An embodiment will be described.

  As shown in FIG. 1, the trust relationship strength estimation apparatus according to the first embodiment estimates the strength of trust relationships between an input unit 1 that accepts input of various data including product evaluation data and users. A calculation unit 2 and an output unit 3 that outputs an estimation result are provided.

  The input unit 1 is realized by an input device such as a known keyboard, mouse, or storage device.

  The output unit 3 is implemented by a display, a printer, a magnetic disk, or the like.

  The calculation unit 2 is a computer having a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) storing a program for executing a trust relationship strength estimation processing routine described later. It is configured and functionally configured as follows. The calculation unit 2 includes a data storage unit 21, a data reading unit 22, a parameter initial value setting unit 23, a confidence relationship strength estimation unit 24, and a convergence determination unit 25. The data storage unit 21 is an example of a storage unit, and the convergence determination unit 25 is an example of an output unit.

The data storage unit 21 stores product evaluation data S that is a set of information (u, i, r _ui ) representing the evaluation value r _ui given to the product i by the user u as a set of observation data. Further, the data storage unit 21 stores trust relationship data G composed of a matrix having 0 and 1 as elements indicating whether or not there is a trust relationship between users in the user set U. The trust relationship data G is obtained in advance as observation data and stored in the data storage unit 21. The data storage unit 21 stores parameters α, β, and γ determined in advance.

  The data reading unit 22 acquires product evaluation data S, trust relationship data G, and parameters α, β, and γ from the data storage unit 21. Note that the data reading unit 22 may obtain the optimum values of the parameters α and β instead of obtaining the data from the data storage unit 21.

The parameter initial value setting unit 23 is an initial value of the parameters s _uv , p _uk , q _ki (1 ≦ u, v ≦ M, 1 ≦ i ≦ N, 1 ≦ k ≦ K) estimated by the trust relationship strength estimation unit 24. Set. For example, a random number is set as the initial value of each parameter.

The trust relationship strength estimation unit 24 uses the product evaluation data S, trust relationship data G, and parameters α, β, and γ to calculate the values of the parameters s _uv , p _uk , and q _ki according to the above equation (3). Update.

Convergence determination unit 25, the parameter s _uv, p _uk, determines whether q _ki has converged, when not converged, the parameter s _uv according trusted intensity estimating unit 24, p _uk, q Repeat the update of _ki . Convergence determination unit 25, when it is determined that convergence has been achieved, the parameter s _uv was last updated, p _uk, the value of q _ki, parameter s _uv that minimizes the evaluation function of the equation (2), It outputs to the output part 3 as _puk and _qki .

As a method for determining whether or not the convergence determination unit 25 has converged, the above equation (2) is calculated using the parameters s _uv , p _uk , and q _ki updated by the trust relationship strength estimation unit 24, and the above (2 A method for determining that the value has converged may be used when the value calculated by the expression) is equal to or less than a predetermined threshold value. As another method, when the difference between the parameters s _uv , p _uk , and q _ki before and after the update in the trust relationship strength estimation unit 24 is equal to or less than a predetermined threshold, it may be determined that the convergence has occurred. For example, if the parameter before update is s _uv , the parameter after update is ^ s _uv , and the threshold is θ, it is determined that convergence has been satisfied when | ^ s _uv −s _uv | <θ, In the case, it is determined that it has not converged.

  Alternatively, the number of repetitions of the trust relationship strength estimation unit 24 may be counted, and when the predetermined number of repetitions is exceeded, it may be determined that the process has converged and the process may be terminated.

The estimation results of the final parameters s _uv , p _uk , and q _ki are output from the output unit 3 to the user.

<Operation of the trust relationship strength estimation device>

  Next, the operation of the trust relationship strength estimation apparatus according to the first embodiment will be described. First, product evaluation data S evaluated by a plurality of users for a plurality of products is input to the trust relationship strength estimation apparatus via the input unit 1 as a set of observation data and stored in the data storage unit 21. In addition, trust relationship data G indicating whether each of a plurality of users trust other users is input to the trust relationship strength estimation apparatus via the input unit 1 and stored in the data storage unit 21. Various constants are input to the trust relationship strength estimation apparatus via the input unit 1 and stored in the data storage unit 21.

  Then, a trust relationship strength estimation processing routine shown in FIG. 2 is executed in the calculation unit 2 of the trust relationship strength estimation device.

First, in step 100, the product evaluation data S, the trust relationship data G, and various parameters stored in the data storage unit 21 are read and acquired. In step 102, initial values of parameters s _uv , p _uk , q _ki (1 ≦ u, v ≦ M, 1 ≦ i ≦ N, 1 ≦ k ≦ K) are set using random numbers.

In the next step 104, the initial value is set in step 102 according to the above equation (3) using the evaluation data S, the trust relationship data G, and the parameters α, β, and γ read in step 100. each parameter s _uv was, p _uk, the value of q _ki, or the parameters s _uv updated in the previous step 104, p _uk, updates the value of q _ki.

In step 106, as a predetermined convergence condition, it is determined whether or not the value calculated by the above equation (2) is equal to or less than a threshold value. If the value calculated by the above equation (2) is larger than the threshold value, Determines that the predetermined convergence condition is not satisfied, returns to step 104, and repeats the process of step 104. On the other hand, if the value calculated by the above equation (2) is equal to or smaller than the threshold value, it is determined that a predetermined convergence condition is satisfied, and the parameter s _uv finally updated in step 106 is determined in step 108. , P _uk , q _ki are output by the output unit 3, and the trust relationship strength estimation processing routine is terminated.

As described above, according to the trust relationship strength estimation apparatus according to the first embodiment, the evaluation data of a product composed of a set of (user, product, evaluation value) is obtained like data on a review site of the product. obtained by using the evaluation value estimated on the basis of the parameters _{s uv, p uk, q ki} , and evaluation values of the other users, the evaluation function using the error between the evaluation values of the observed data, the parameter s _uv , P _uk , q _ki can be optimized to estimate the strength of the trust relationship between users.

  In addition, by virtualizing the users that each user trusts and adding the evaluation data of the virtual users to the set of observation data, even if there is a user that no user trusts, The strength of the trust relationship can be estimated.

  Next, a second embodiment will be described. In the second embodiment, a case where the present invention is applied to a product evaluation value prediction apparatus that predicts a user's evaluation value for a product will be described as an example. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

The second embodiment is different from the first embodiment in that a user's evaluation value for a product is predicted using the estimated parameters s _uv , p _uk , and q _ki .

  As shown in FIG. 3, an input unit 1, a computing unit 202 that estimates the strength of a trust relationship between users, predicts a user's evaluation value, and an output unit 3 that outputs a prediction result are provided. .

  The calculation unit 202 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a reliability relationship strength estimation processing routine and a product evaluation value prediction processing routine described later. As shown in FIG. The calculation unit 202 includes a data storage unit 21, a data reading unit 22, a parameter initial value setting unit 23, a confidence relationship strength estimation unit 24, a convergence determination unit 25, an estimation result storage unit 220, and a product evaluation prediction unit 221. . The estimation result storage unit 220 is an example of a parameter storage unit.

Convergence determination unit 25, the parameter s _uv, p _uk, it is determined whether or not converged q _ki, if it is determined that convergence has been achieved, the last updated parameters s _uv, p _uk, the q _ki The value is stored in the estimation result storage unit 220 as parameters s _uv , p _uk , q _ki that minimize the evaluation function.

The estimation result storage unit 220 stores estimated values of the parameters s _uv , p _uk , q _ki (1 ≦ u, v ≦ M, 1 ≦ i ≦ N, 1 ≦ k ≦ K).

The product evaluation prediction unit 221 is based on the product evaluation data stored in the data storage unit 21 and the estimated values of the parameters s _uv , p _uk , q _ki stored in the estimation result storage unit 220 ( 1) The evaluation value ^ r _ui of the prediction target user u for the prediction target product i is predicted according to the equation (1).

  The evaluation value prediction result is output from the output unit 3 to the user.

<Operation of the product evaluation value prediction device>
Next, the operation of the product evaluation value prediction apparatus according to the second embodiment will be described.

  First, as in the first embodiment, the product evaluation data S, the trust relationship data G, and various constants are input to the product evaluation value prediction device via the input unit 1 and stored in the data storage unit 21. Stored.

Then, in the calculation unit 202 of the product evaluation value prediction apparatus, the trust relationship strength estimation processing routine shown in FIG. 2 is executed, and the estimated values of the parameters s _uv , p _uk , and q _ki are stored in the estimation result storage unit 220. The

Also, the index u of the user to be predicted and the index i of the product for which the evaluation value r _ui is not obtained from the user set U and the product set I in the product evaluation data S via the input unit 1 are the product When input to the evaluation value prediction apparatus, the calculation unit 202 executes a product evaluation value prediction processing routine shown in FIG.

First, in step 250, the input index u of the prediction target user and the index i of the product are acquired. In step 252, the evaluation data relating to the product i is read from the data storage unit 21, and the parameters p _uk , q _ki relating to the prediction target user u and the product i from the estimation result storage unit 220 (1 ≦ k ≦ K). Then, each parameter s _uv (1 ≦ v ≦ M, v ≠ u) and parameter α relating to the user u to be predicted is read.

In the next step 254, the predicted value ^ r _ui of the evaluation value of the user u for the product i is calculated according to the above equation (1) based on the various data read in the step 252.

Then, in step 256, the prediction value ^ r _ui of the evaluation value calculated in step 254 is output by the output unit 3, and the product evaluation value prediction processing routine is terminated.

<Experimental example>
Next, the result of an experiment using the evaluation value prediction method proposed in this embodiment will be described. In the experiment, the experiment was performed using product evaluation data collected from Epines of the product review site.

  As a comparison method, evaluation values were predicted using a standard matrix decomposition method expressed by the following equation (4).

That is, the parameters p _uk and q _ki were estimated using the evaluation function represented by the following equation (5), and the evaluation value of the user u for the product i was predicted using the above equation (4).

  Moreover, the mean square error (RMSE) represented by the following formula (6) was used as an evaluation standard for the experimental results.

  Further, in the experiment, in order to investigate how the value of the parameter α affects the result, the experiment was performed by manually setting the value of the parameter α. For other parameters, the learning metaparameter γ = 0.001, the parameter β = 0.01, and the number of potential features K = 20.

  As a set of observation data, each parameter was learned using 80% of data collected from the review site, and when the evaluation value was predicted using the learned parameter, a result as shown in FIG. 5 was obtained. In FIG. 5, Trust represents the result of the experiment using the evaluation value prediction method proposed in the present embodiment, and MF represents the result of the experiment using the standard matrix decomposition method. Mean represents a result when an average value of evaluation values given to the products is set as a predicted value.

  When the evaluation value prediction method proposed in this embodiment is used, a lower RMSE is achieved than when the standard matrix decomposition method is used. It was confirmed that the prediction of the evaluation value becomes more accurate by estimating the strength of the trust relationship between users and using it for prediction of the evaluation value.

As described above, according to the evaluation value predicting apparatus according to the second embodiment, the evaluation value estimated based on the parameters s _uv , p _uk , q _ki , and the evaluation values of other users, and the observation The strength s _uv of the trust relationship between users can be estimated by optimizing the parameters s _uv , p _uk , and q _ki using an evaluation function that uses an error from the evaluation value of the data. Also, the user's evaluation value for the prediction target item can be accurately predicted using the parameters s _uv , p _uk , q _ki and the evaluation value observation data.

  In addition, the strength of the trust relationship between users can be estimated, and this estimation result can be used to predict what kind of evaluation a user has for a certain product. A product with a high evaluation value can be recommended.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, you may use the evaluation value with respect to items (a web page, a photograph, etc.) other than goods.

  Moreover, although the above-mentioned reliability relationship strength estimation apparatus and evaluation value prediction apparatus have a computer system inside, if the “computer system” uses a WWW system, the homepage provision environment (or Display environment).

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 1 Input part 2,202 Operation part 3 Output part 21 Data storage part 22 Data reading part 23 Parameter initial value setting part 24 Reliability relation strength estimation part 25 Convergence determination part 220 Estimation result storage part 221 Product evaluation prediction part

Claims (8)

Storage means for storing, for a plurality of users and a plurality of items, a set of observation data including user information indicating the users, item information indicating the items, and evaluation values evaluated by the users for the items; ,
An initial value for setting an initial value of each of a user feature parameter that represents each feature of the plurality of users, an item feature parameter that represents each feature of the plurality of items, and a strength parameter that represents the strength of the trust relationship between the users. Value setting means;
For each of the observation data, the initial value or the last updated value of the user feature parameter of the user information of the observation data, and the initial value or the last updated value of the item feature parameter of the item information of the observation data A value output from a function that takes a value as input, an initial value or a previously updated value of the intensity parameter between the user indicated by the user information and another user, and the observation data obtained from the set of observation data By the evaluation function using the error between the evaluation value calculated by adding the value output from the function that receives the evaluation value of each of the other users for the item information, and the evaluation value of the observation data, Each of the user characteristic parameters of each of the plurality of users, the item characteristic parameters of each of the plurality of items, and each between the users And estimating means for updating the value of each parameter the value of said each respective value is calculated, and the calculated between the intensity parameter,
The user characteristic parameter by said estimating means, the item feature parameters, and by repeating the update of the strength parameters, the set of pre-Symbol observation data, the user characteristic parameter, the item characteristic parameter, and the intensity parameter An output means for outputting the user feature parameter, the item feature parameter, and the intensity parameter when optimized,
A reliability relationship strength estimation device including :
The user feature parameter of the user information is a degree of association between the user information and a potential feature k,
The item information parameter of the item information is a degree of association between the item information and the potential feature k.
Trust relationship strength estimation device. The calculated evaluation value of the user information u for the item information i is:
For the plurality of potential features k, a product obtained by multiplying the degree of association between the user information u and the potential feature k and the degree of association between the potential feature k and the item information i is added. And the total value
For each of the user information v of the user trusted by the user indicated by the user information u obtained in advance, the strength parameter between the user information u and the user information v, and the item information i The total value obtained by adding the products of the evaluation values of the user information v is divided by the total value obtained by adding the intensity parameters between the user information u and the user information v for each of the user information v. Quotient,
The reliability relationship strength estimation apparatus according to claim 1, which is calculated by weighted addition. The set of observation data includes a plurality of observation data including an average value of evaluation values for the user information, the item information, and the item information, which are generated for each of the item information, indicating a virtual user,
The trust relationship strength estimation apparatus according to claim 2, wherein each of the plurality of users trusts the virtual user. The reliability relationship strength estimation device according to any one of claims 1 to 3,
Parameter storage means for storing the user feature parameter, the item feature parameter, and the strength parameter output by the trust relationship strength estimation device;
The evaluation value of the prediction target user among the plurality of users with respect to the prediction target item of the plurality of items, the user feature parameter of the user information stored in the parameter storage unit, the item feature parameter And item evaluation prediction means for predicting using the intensity parameter,
Evaluation value prediction apparatus including The item evaluation prediction means includes:
The evaluation value of the user information u for the item information i is
For a plurality of potential features k, a product obtained by multiplying the degree of association between the user information u and the potential feature k and the degree of association between the potential feature k and the item information i is added. The total value,
For each of the user information v of the user trusted by the user indicated by the user information u obtained in advance, the strength parameter between the user information u and the user information v, and the item information i The total value obtained by adding the products of the evaluation values of the user information v is divided by the total value obtained by adding the intensity parameters between the user information u and the user information v for each of the user information v. The evaluation value prediction apparatus according to claim 4, wherein the quotient and are calculated by weighted addition. Storage means for storing, for a plurality of users and a plurality of items, a set of observation data including user information indicating the users, item information indicating the items, and evaluation values evaluated by the users for the items; A confidence relationship strength estimation method in a confidence relationship strength estimation device including an initial value setting means, an estimation means, and an output means,
The trust relationship strength estimation device is:
Each of a user feature parameter representing each feature of the plurality of users, an item feature parameter representing each feature of the plurality of items, and an intensity parameter representing the strength of the trust relationship between the users by the initial value setting means. A step of setting an initial value of
An initial value of the user characteristic parameter of the user information of the observation data or a previously updated value and an initial value of the item characteristic parameter of the item information of the observation data for each of the observation data by the estimation unit Or a value output from a function that receives a previously updated value, an initial value of the intensity parameter between the user indicated by the user information and another user, or a previously updated value, and the observation data The error between the evaluation value calculated by adding the values output from the function that receives the evaluation values of the other users for the item information obtained from the set and the evaluation value of the observation data is used. The user characteristic parameter of each of the plurality of users and the item characteristic parameter of each of the plurality of items The steps of the calculating each value of the intensity parameter of each among users, updates the value of each parameter calculated the respective values,
By said output means, the user characteristic parameter by said estimating means, the item feature parameters, and by repeating the update of the strength parameters, the set of pre-Symbol observation data, the user characteristic parameter, the item characteristic parameter, And outputting the user feature parameter, the item feature parameter, and the strength parameter when the strength parameter is optimized;
A reliability relationship strength estimation method characterized in that it is executed including:
The user feature parameter of the user information is a degree of association between the user information and a potential feature k,
The item information parameter of the item information is a degree of association between the item information and the potential feature k.
Trust relationship strength estimation method . The reliability relationship strength estimation device according to any one of claims 1 to 3, and a parameter storage that stores the user feature parameter, the item feature parameter, and the strength parameter output by the trust relationship strength estimation device. An evaluation value prediction method in an evaluation value prediction apparatus including means and item evaluation prediction means,
The evaluation value prediction apparatus
Each of a user feature parameter representing each feature of the plurality of users, an item feature parameter representing each feature of the plurality of items, and an intensity parameter representing the strength of the trust relationship between the users by the initial value setting means. A step of setting an initial value of
An initial value of the user characteristic parameter of the user information of the observation data or a previously updated value and an initial value of the item characteristic parameter of the item information of the observation data for each of the observation data by the estimation unit Or a value output from a function that receives a previously updated value, an initial value of the intensity parameter between the user indicated by the user information and another user, or a previously updated value, and the observation data The error between the evaluation value calculated by adding the values output from the function that receives the evaluation values of the other users for the item information obtained from the set and the evaluation value of the observation data is used. The user characteristic parameter of each of the plurality of users and the item characteristic parameter of each of the plurality of items The steps of the calculating each value of the intensity parameter of each among users, updates the value of each parameter calculated the respective values,
By said output means, the user characteristic parameter by said estimating means, the item feature parameters, and by repeating the update of the strength parameters, the set of pre-Symbol observation data, the user characteristic parameter, the item characteristic parameter, And outputting the user feature parameter, the item feature parameter, and the strength parameter when the strength parameter is optimized;
The user of the user information stored in the parameter storage means stores the evaluation value of the prediction target user of the plurality of users for the prediction target item of the plurality of items by the item evaluation prediction means. Predicting using the feature parameter, the item feature parameter, and the intensity parameter;
An evaluation value prediction method characterized in that it is executed including:
The user feature parameter of the user information is a degree of association between the user information and a potential feature k,
The item information parameter of the item information is a degree of association between the item information and the potential feature k.
Evaluation value prediction method .   A program for causing a computer to function as each means of the reliability relationship strength estimation apparatus according to any one of claims 1 to 3 or each means of the evaluation value prediction apparatus according to claim 4 or 5.

Images