JP2007323315A - Cooperative filtering method, cooperative filtering device, cooperative filtering program and recording medium with the same program recorded thereon - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a cooperative filtering method, a cooperative filtering device, a cooperative filtering program and a recording medium with the program recorded thereon for predicting the unevaluated scores of a user with higher precision than a conventional manner. <P>SOLUTION: In the cooperative filtering method for predicting the scores of the unevaluated items of a certain user by using an evaluation history including scores evaluated by a plurality of users with a discrete value for a plurality of items, an input data reading part 221 reads a matrix R as input data, a prediction score initial value calculation part 222 substitutes an initial value into the unevaluated scores of the matrix R, a prediction score updating part 223 predicts the unevaluated scores by decreasing the KL divergence of an evaluated score group and unevaluated score group, and a prediction score writing part 224 writes the acquired prediction scores in a storage means 4. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a collaborative filtering method, a collaborative filtering device, a collaborative filtering program, and a recording medium recording the program, which can predict a score that has not been evaluated by a user with higher prediction accuracy than before.

  Recommendation services that select and recommend items that are predicted to give high ratings from users are used mainly in the field of e-commerce (for example, products) , Rental video, TV program, or restaurant recommendation system). A technique for predicting an unrated score of an active user by using not only the score history but also the score history of other users for a recommendation target user (hereinafter referred to as an active user) is called collaborative filtering.

  By preferentially recommending items with high predictive scores through collaborative filtering, it is possible to reduce the burden of item selection for active users, and at the same time, it is possible to present items that the active user did not know Become. In addition, on the item provider side, activation of item sales activities and efficiency can be promoted by recommending appropriate items to the user.

  Collaborative filtering is based on the score history of many users (for example, on the history data of scores given by users to items as shown in Table 1, a user with a user ID of 196 scores an item with an item ID of 242) As a result, the score for the unrated item of the active user is predicted. Then, an item with a high predicted score value is preferentially recommended. Therefore, in the recommendation system, how to predict the score value of the unscored score from the score history with high accuracy, that is, the realization of the highly accurate score prediction method is an important issue.

  Conventional score prediction methods are largely divided into two approaches: (i) a method based on the nearest neighbor method (for example, see Non-Patent Document 1) and (ii) a method based on a probability model (for example, see Non-Patent Document 2). Can be separated. (I) searches a plurality of users having a score history similar to the score history of the active user, and predicts the score of the unscored item of the active user using the score history of the searched user. In (ii), first, a probability model expressing the user's score history is assumed, and the model is learned using the actual score history. Thereafter, the score of the unscored item of the active user is predicted using the learned model.

  In any of the approaches, it can be said that the prediction is performed using some similarity assumption between users. In other words, (i) uses the assumption that if the grades of the scored items common to the active user and the user to be searched are similar to each other, the unscored parts are also similar.

  In (ii), the user belongs to several classes with similar preferences (in (ii), it is assumed that such a class exists. Hereinafter, this group is referred to as a latent class). The assumption is that the item scores of users belonging to the same latent class follow the same probability distribution within that latent class.

  Similarity by any of these methods can be said to be “local similarity” in the sense that only part of the information of the user or item is used when predicting an unrated score. That is, there is a problem that there is a high possibility that high prediction accuracy cannot be obtained because an unrated score is predicted without capturing the global characteristics of the entire score.

P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl.Grouplens: An Open Architecture for Collaborative Filtering of Netnews.In Proceedings of ACM CSCW1994, pages 175-186, October 1994. T. Hofmann, "Latent Semantic Models for Collaborative Filtering", ACM Trans. Information Systems, 22 (1): 89-115, January 2004.

  The present invention has been made in view of the above-mentioned problems, and is highly accurate by a score prediction method that predicts unscored scores based on “global similarity” using all information about users and items. Filtering method, collaborative filtering device, collaborative filtering program, and recording medium on which the program is recorded are provided.

  In order to solve the above-mentioned problem, the collaborative filtering method according to claim 1 is characterized in that a score of an unevaluated item of a certain user is obtained by using a score history including scores evaluated by a large number of users with respect to a large number of items. Is a collaborative filtering method for predicting an experience score distribution calculating step for outputting an experience score distribution indicating a distribution of evaluated scores inputted as a score history, and substituting a predetermined initial prediction value for an unrated score. Predictive score distribution calculation step that outputs predicted score distribution, dissimilarity calculation step that outputs dissimilarity between experience score distribution and predicted score distribution, and nonlinear optimization calculation so that dissimilarity is reduced And a score estimation step of sequentially estimating and outputting unrated ratings.

  According to such a procedure, in the experience score distribution calculation step, an experience score distribution indicating the distribution of evaluated scores is output, and in the predicted score distribution calculation step, a predetermined predicted initial value is substituted for an unrated score. The score distribution is output, the dissimilarity between the experience score distribution and the predicted score distribution is output in the dissimilarity calculation step, and the dissimilarity is reduced in the score estimation step, that is, the predicted score distribution is Non-evaluated scores are estimated by performing nonlinear optimization calculations so as to approximate the empirical score distribution.

  Further, the collaborative filtering method according to claim 2 is the collaborative filtering method according to claim 1, wherein the experience score distribution calculating step calculates the frequency of the score of discrete values over the items already scored for each user. , The step of outputting the experience score distribution for each user, the step of outputting the frequency of discrete score across the users who have already been scored for each item, the step of outputting the experience score distribution for each item, and all the score data Calculating a discrete score and outputting an experience score distribution for each user and item.

  According to this procedure, in the experience score distribution calculation step, (1) the frequency of discrete score is calculated over the items already scored for each user, and the experience score distribution for each user is output. Outputs the frequency of discrete score across the users who have already been scored, outputs the empirical score distribution for each item, and (3) calculates the discrete score for all score data, for each user and item Output experience score distribution.

  Further, the collaborative filtering method according to claim 3 is the collaborative filtering method according to claim 1, wherein the predictive score distribution calculating step calculates the frequency of the score of discrete values over the items already scored for each user. , A step of outputting a predicted score distribution for each user, a step of calculating a frequency of a discrete score across users who have already been scored for each item, and a step of outputting a predicted score distribution for each item, and for all score data And calculating a discrete score and outputting a predicted score distribution for each user and item.

  According to such a procedure, in the predicted score distribution calculating step, (1) the frequency of discrete score is calculated over the items already scored for each user, and the predicted score distribution for each user is output. (2) Item The frequency of discrete score is calculated for each scored user, and the predicted score distribution for each item is output. (3) The discrete score is calculated for all score data, Output the predicted score distribution.

  The collaborative filtering method according to claim 4 is the collaborative filtering method according to claim 1, wherein the dissimilarity calculation step calculates the KL divergence between the experience score distribution and the predicted score distribution. Is calculated.

  According to this procedure, the dissimilarity is calculated by calculating the KL divergence between the experience score distribution and the predicted score distribution in the dissimilarity calculation step.

  Further, the collaborative filtering method according to claim 5 is the collaborative filtering method according to claim 1, wherein the score estimation step is optimal so that the sum of the amount of the Cullback information between the experience score distribution and the predicted score distribution is minimized. It is characterized by performing calculus calculation.

  According to such a procedure, in the score estimation step, the optimization calculation is performed so that the total sum of the Cullback information amounts of the experience score distribution and the predicted score distribution is minimized.

  Further, the collaborative filtering device according to claim 6 predicts the score of an unscored item of a certain user by using a score history including scores evaluated by a large number of users with discrete values for a large number of items. An equipment that outputs an empirical score distribution indicating the distribution of evaluated scores based on the score history data, and a predicted score distribution by substituting a predetermined predicted initial value for an unscored score. Predictive score distribution calculator to output, dissimilarity calculator to calculate dissimilarity between experience score distribution and predicted score distribution, and non-evaluation by performing nonlinear optimization calculation so that dissimilarity is reduced And a score estimation unit that sequentially estimates the score.

  According to this configuration, the empirical score distribution calculation unit outputs an empirical score distribution indicating the distribution of evaluated scores, and the predicted score distribution calculation unit substitutes a predetermined predicted initial value for an unevaluated score. The score distribution is output, the dissimilarity calculation unit outputs the dissimilarity between the experience score distribution and the predicted score distribution, and the score estimation unit reduces the dissimilarity, that is, the predicted score distribution is Non-evaluated scores are estimated by performing nonlinear optimization calculations so as to approximate the empirical score distribution.

  A collaborative filtering program according to claim 7 causes a computer to execute the collaborative filtering method according to claim 1. With this configuration, the computer can realize the method according to claim 1. That is, according to the collaborative filtering program that executes the collaborative filtering method based on such a score prediction method, the same function as that of the collaborative filtering device based on the score prediction method that simultaneously predicts all scores of the prediction target based on the global similarity. Can be executed by a computer.

  A recording medium according to an eighth aspect stores the collaborative filtering program according to the seventh aspect. With this configuration, it is possible to cause a desired computer to execute the method described in claim 1. That is, according to the recording medium storing the collaborative filtering program for executing the collaborative filtering method based on such a score prediction method, the collaborative filtering based on the score prediction method that simultaneously predicts all the scores of the prediction target based on the global similarity. A program for causing a computer to execute the same function as the apparatus can be recorded in a recording medium.

  According to the present invention, the global similarity including more information than before, that is, based on the similarity between the experience peripheral score distribution regarding the scored data and the prediction score distribution regarding the unrated score data to be predicted with high accuracy. Score prediction can be performed.

The present invention has been developed to solve the above-described problems, and provides a score prediction method for predicting a score to be predicted based on global similarity.
In the score prediction method according to the present invention, first, the peripheral score distribution for each user, the peripheral score distribution for each item, and the peripheral score distribution of all scores are calculated from the scored data. These are obtained by counting the frequency from the actual observation values. These score distributions are referred to as experience marginal score distributions.

  In the present description, among a plurality of score distributions, a peripheral score distribution based on a predetermined condition is referred to as a peripheral score distribution or simply a score distribution. Further, “score” means not only an evaluated point but also an evaluation (adding a score).

Here, the peripheral score distribution means a score distribution from a certain viewpoint. For example, a score history matrix (table) as shown in FIG. 1 is used, with positive values representing score values and 0 representing unscored values. For example, it means that the user 1 gives a score 2, 3, 1 to the items 1, 3, 4 respectively, and the items 2, 5 do not give a score. And experience near score distribution for each user with respect to this data, scores distribution in the row direction of the matrix of FIG. _{2 (a) (P ~ i} , i = 1, ..., 6) refers to a empirical peripheral score distribution for each item Indicates the score distribution (Q ~ _j , j = 1,..., 5) in the column direction of the matrix of FIG. 2 (a). The score distribution of all scores is the score distribution of the entire matrix of FIG. 2 (a). (S ~). For example, P ~ _1, the probability scores 1, 2 and 3 indicate that all is 1/3.

Then, the peripheral score distribution is expressed as a function of the predicted value for the unrated score data to be predicted. For example, P _i , Q _j , and S in FIG. 2B represent an example of a predicted peripheral score distribution calculated when a predicted value in an unrated score is substituted. Then, the dissimilarity between the two distributions (for example, KL divergence) is set so that the corresponding marginal score distributions (P ~ _i and _Pi , Q ~ _j and _Qj , S ~ and S) are as similar as possible. The predicted value of the unrated score is obtained by minimizing.

  The prediction target score group contributes to the peripheral score distribution for each user, the peripheral score distribution for each item, and the score distribution of all the scores, so that the prediction target scores are directly or indirectly related to each other. A prediction based on “sex” will be made. Therefore, improvement in prediction accuracy can be expected as compared with the prior art. A similar function can be executed by a computer. Furthermore, according to the recording medium storing the score prediction program for executing such a score prediction method, the program for causing the computer to execute the same function as the device for simultaneously predicting all the scores to be predicted based on the global similarity Can be recorded in a recording medium.

Next, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
FIG. 3 is a block diagram showing the collaborative filtering device 1 according to the embodiment of the present invention.
As shown in FIG. 3, the collaborative filtering device 1 includes a calculation unit 2, an input unit 3, a storage unit 4, an output unit 5, and a bus line 11 that connects these units.

  The computing means 2 is a central part of a computer composed of a CPU, a RAM and the like. The computing means 2 includes a preprocessing unit 21, a score prediction unit 22, and a memory 23. The calculation means 2 reads the preprocessing program 41 and the score prediction program 42 from the storage means 4, stores them in the memory 23, and executes them, thereby realizing the preprocessing section 21 and the score prediction section 22.

  The input means 3 is composed of a keyboard, a disk drive device, etc. (both not shown). The rating history is input via the input unit 3 and stored in the storage unit 4.

  The storage means 4 is composed of a hard disk device or the like. The storage unit 4 can store a preprocessing program 41 and a score prediction program 42. The storage unit 4 includes a score history 43, input data 44, and a predicted score 45.

Here, as shown in Table 1, the score history is a history of user ID, item ID, score, time stamp, and the like. The rating ri _{, j} for the user ID i (i = 1, 2,..., N) and the item ID j (j = 1, 2,..., N) is a discrete value {1, 2,. } The higher the score r _{i, j,} the better the evaluation.

  The output means 5 is, for example, a graphic board and a monitor connected thereto, and displays items recommended for the active user.

FIG. 4 is a block diagram showing the preprocessing unit 21 of the present embodiment.
The preprocessing unit 21 includes a grade history reading unit 211 and an input data writing unit 212.

  The score history reading unit 211 reads the user ID, item ID, and score from the score history 43.

Then, the input data writing unit 212 stores the user ID, the item ID, and the grade in the input data 44. As shown in FIG. 1, the input data 44 is a matrix having user IDs as rows, item IDs as columns, and elements as grades. Assume that “0” is included in the score that is not rated by the user. Hereinafter, a matrix of input data is denoted as R. The (i, j) element of the matrix R represents the rating r _{i, j} .

Here, for simplification of description, the following variables are defined for the matrix R.
A set of r _{i, j} , jεJ _i 1 with user ID i and graded is denoted as B _i 1, and a set of unrated scores r _{i, j} , jεJ _i 0 is denoted as B _i 0. Here, J _i 1 (J _i 0) represents an item index set (not rated) by a user whose user ID is i. For convenience, the score set for the item ID i in the matrix R is denoted as B _j 1 and the unrated score set is denoted as B _j 0. That is, in the matrix R, the (un) score set of the i-th row is B _i 1 (B _i 0), and the (un) score set of the j-th column is B _j 1 (B _j 0). Further, if i and j are not particularly considered, the score set that has already been scored in the matrix R is denoted as B ¹ _* , and the score set that has not been scored is denoted as B ⁰ _* .

FIG. 5 is a block diagram of the score prediction unit 22 according to the present embodiment.
The score predicting unit 22 includes an input data reading unit 221, a predicted score initial value calculating unit 222, a predicted score updating unit 223, and a predicted score writing unit 224.

  The input data reading unit 221 reads the matrix R from the input data 44.

The predicted score initial value calculation unit 222 is an unrated score for all the unscored scores {ri _, jεB ⁰ _* } to be predicted among the elements of the matrix R read by the input data reading unit 221. Some numerical value is substituted as an initial value instead of 0 indicating the presence. For example, instead of 0, 3 can be substituted for {ri _, jεB ⁰ _* }. It is also possible to substitute a value calculated by some means. The score assigned with the initial value is expressed as {r ⁽⁰⁾ _{i, j} ∈ B ⁰ _* }.
A real value can be substituted as an initial value. In the following description, the score r _{i, j} is assumed to be a real value.

  The predicted score update unit 223 calculates an unscored score set that minimizes a certain objective function based on the initial value for the unscored score set created by the predicted score initial value calculation unit 222.

More specifically, first, the score history data consisting of N users and M items (N × M matrix R, i-th and j-th elements r _{i, j} {0, 1,... Represents the score of item j. The maximum value of the score is V, and 0 represents an unrated score).

The experience peripheral score distribution P ~ _i , i = 1, 2,..., N in the row direction (for each user) of the matrix R, and the predicted peripheral score distribution P ~ _i , i = 1, 2,.
· Experience peripheral score distribution Q ~ _i , i = 1, 2, ..., N in the row direction (for each item) of matrix R, and predicted peripheral score distribution Q ~ _i , i = 1, 2, ..., N.
The empirical score distribution S ~ of the entire matrix R and the predicted peripheral score distribution S.

FIG. 2 shows an example where N = 6 and M = 5.
For example, user i's experience peripheral score distribution P ~ _i uses a score set B _i 1 {ri _{, j} ; ri _{, j} > 0, j = 1,. Calculate with the following formula.

Here, δ (x) represents a delta function, and takes a value of 1 when x is true, and 0 otherwise. That is, P _i (k) corresponds to the frequency ratio of the score k to the other scores l ≠ k. Obviously, Σ _k P ~ _i (k) = 1.

On the other hand, the predicted peripheral score distribution P _i related to unscored scores is calculated by the following equation using the initial value {r ⁽⁰⁾ _{i, j} ∈ B _i 0} created by the predicted score initial value calculation unit 222.

B _i 0 represents a score set B _i 0 = {r _{i, j} | r _{i, j} = 0, j = 1,... σ (> 0) is a smoothing parameter. The larger σ is, the more effective it is to reduce the deviation between the predicted value and the true value. In the limit of σ → 0, Equation (2) matches the form of Equation (1).

Next, using the peripheral score distribution created by the above calculation, the initial value {r ⁽⁰⁾ _{i, j} ∈ B ⁰ _* } created by the predicted score initial value calculation unit 222 and the scored score { The dissimilarity of each peripheral distribution of r _{i, j} ∈ B ¹ _* } is calculated. A known measure of KL divergence can be applied as the dissimilarity. For example, the KL divergence of the marginal distribution regarding the user i is calculated by the following equation.

Similarly for Q _j and S, the dissimilarities KL ( _Q˜j || _Qj ) and KL ( _S˜j || _Sj ) are calculated by the following equations.

By the above calculation, the initial value {r ⁽⁰⁾ _{i, j} ∈ B ⁰ _* } created by the predicted score initial value calculation unit 222 and the score already obtained {ri _{, j} ∈ B ¹ _* } are summarized. The similarity is calculated by the objective function J {r ⁽⁰⁾ _{i, j} ∈ B ⁰ _* } shown below.

After calculating the objective function based on global similarities, subsequently, the objective function _{^{J {r (0) i,}} j ∈B 0 *} score set of non scores minimize _{{r i, j | r i} , Find _j ∈ B ⁰ _* }. A score set of unscores obtained as a result of minimizing the objective function is set as {r ^(f) _{i, j} | r ^(f) _{i, j} ∈B ⁰ _* }. Here, as means for minimizing the objective function, nonlinear optimization means such as a quasi-Newton method can be applied.
Finally, the unrated score set {r ^(f) _{i, j} | r ^(f) _{i, j} ∈ B ⁰ _* } is stored in the predicted score 45.

FIG. 6 is a data flow diagram of the collaborative filtering device 1.
When the rating history 43 is input, the preprocessing unit 21 performs the above-described preprocessing, generates input data, and temporarily stores it in the storage unit 4 (see FIG. 3). The preprocessing unit 21 sends this input data to the score prediction unit 22. The score prediction unit 22 performs prediction related to the score and outputs a prediction score 45, and the prediction score 45 is accumulated in the storage unit 4.

  In the present invention, a measure other than KL divergence (equation (3), equation (4), equation (5)) may be used as a measure for measuring the similarity between peripheral score distributions that are the basis of global similarity. In addition, by minimizing the objective function derived in the same manner, it is possible to predict a score that has not been scored.In addition, the present invention provides various optimization methods for minimizing the objective function (equation (6)). The method is applicable.

The following three methods were applied to the actual data.
· Score prediction method in the present invention (hereinafter referred to as “invention method”)
A method based on the nearest neighbor method (hereinafter referred to as “conventional method A”)
A method based on a probability model (hereinafter referred to as “conventional method B”)
For the conventional method A, refer to Non-Patent Document 1, and for the conventional method B, refer to Non-Patent Document 2.

In the conventional method A, first, a “similar user” having a score history similar to that of the active user a is searched, and then prediction is performed using the scored scores r _{u, j} of the searched similar user u. Score _{r a,} the predicted value for the _{j r ^ a,} prediction equation for obtaining the _j is given by.

Here, S _a is a set of similar users for the active user a, w _{a, u} is the similarity between the active user a and the user u, and r ￣ _a is the average of the rated ra _{, j} of the active user a Represents each value. The similarity wa _{, u} is calculated by the Pearson correlation coefficient, the cosine similarity, and the like using the score that has already been scored by the active user a and the user u.

Conventional method B is a mixture of “the user is divided into C latent classes z _c (c = 1, 2,..., C) having similar preferences, and the score is a sampling from a normal distribution set for each latent class”. A normal distribution model is assumed. However, the score is standardized for each user (the average of the user is subtracted from the original score and divided by the standard deviation of the user's score), and the score is made real. That is, p (r | a, j) represents the distribution followed by the standardized score r _{a, j} of the active user a, P (z _c | a) represents the probability of belonging to the latent class z _c of the active user a, and the latent class. If the normal distribution followed by the score of z _c is p (r | z _c , j), it can be written as The probability function is shown in upper case letters and the probability density function is shown in lower case letters.

Score _{r a,} the predicted value _{r ^ a} for _j, prediction formula for determining the _j is given by the following equation.

Here, P ^ (z _c | a) represents an estimated value of the belonging probability obtained as a result of learning, and μ ^ ^j _c represents an estimated value of the average value of the normal distribution followed by the score of the item j in the latent class z _c . .

(Data used)
In this embodiment, “MovieLens (ML data)”, which is score data for a movie, is used. The score in this data takes values of 1, ..., 5. As the data to which each method was actually applied, the following data obtained by slightly processing ML data was used.
That is, it is data consisting of 41 users and 50 movies, created by extracting users and items from the original data so that all elements in the matrix R are scored.

(Evaluation scale)
Normalized Mean Absolute Error (NMAE) was used as a scale for evaluating the accuracy of prediction. NMAE is shown below.

  Here, MAE (average absolute prediction error) is given by the following equation.

r ^ _{i, j} represents a predicted score for an unrated score r _{i, j} , and # {} represents the number of elements in the set. The best value of NMAE is 0, and if it is greater than 1, it means worse than predicting the score randomly. E {MAE} can be easily calculated if N, M, and V are known.

(Implementation method)
For the graded scores in the matrix R, 20% is randomly selected for each user as unscored data (test data), and the remaining 80% is designated as graded data (learning data). Five sets of these were prepared, and the average value of NMAE for these five sets was determined to evaluate each method.

(Implementation results)
Table 2 shows NMAE in each prediction method.

  FIG. 7 shows a part of values predicted by the score prediction method according to the present invention. The calculation time in Table 2 is the processing time of a C-language implementation on a computer with an internal clock frequency of 3.60 GHz and a main memory of 2 GB.

  According to the present invention, in particular, in the electronic commerce field, specifically, in a recommendation system such as merchandise sales, rental video, TV program, or restaurant, a user can select from a huge group of items (including services). Items that are expected to give a high evaluation can be selected and recommended more appropriately than before.

  Eventually, by collaborative filtering, it is possible to reduce the burden of item selection for active users by preferentially recommending items with high prediction scores, and at the same time, it is possible to present items that the active user did not know It becomes possible. In addition, on the item provider side, activation of item sales activities and efficiency can be promoted by recommending appropriate items to the user.

It is a figure which shows the example of score distribution expressed with the matrix (table) of the score history. It is a schematic diagram which contrasts the evaluated peripheral score distribution and the unevaluated peripheral score distribution. It is a block diagram which shows the collaborative filtering apparatus which concerns on embodiment of this invention. It is a block diagram which shows the pre-processing part of this embodiment. It is a block diagram of the score estimation part which concerns on this embodiment. It is a data flow figure of a collaborative filtering apparatus. It is a figure which shows a part of value estimated by the score prediction system by this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Collaborative filtering apparatus 2 Computation means 3 Input means 4 Storage means 5 Output means 11 Bus line 21 Pre-processing part 22 Score prediction part 23 Memory 41 Pre-processing program 42 Score prediction program 43 Score history 44 Input data 45 Predictive score 211 Reading score history Unit 212 Input data writing unit 221 Input data reading unit 222 Predictive score initial value calculating unit 223 Predictive score updating unit 224 Predictive score writing unit

Claims (8)

A collaborative filtering method for predicting a score of an unevaluated item of a user using a score history including a score evaluated by a large number of users with a discrete value for a large number of items,
An experience score distribution calculating step for outputting an experience score distribution indicating the distribution of the score that has been input and evaluated, and
A predicted score distribution calculating step for outputting a predicted score distribution by substituting a predetermined predicted initial value for an unrated score,
A dissimilarity calculating step for outputting a dissimilarity between the experience score distribution and the predicted score distribution;
A score estimation step of performing non-linear optimization calculation so as to reduce the dissimilarity, and sequentially estimating and outputting an unevaluated score;
A collaborative filtering method comprising: The experience score distribution calculating step includes:
Calculating a frequency of discrete score across the items evaluated for each user, and outputting an experience score distribution for each user;
Calculating the frequency of discrete score across the evaluated users for each item and outputting an empirical score distribution for each item;
Calculating discrete scores for all score data and outputting an experience score distribution for each of the user and the item;
The collaborative filtering method according to claim 1, further comprising: The predicted score distribution calculating step includes:
Calculating a frequency of discrete score across items evaluated for each user, and outputting a predicted score distribution for each user;
Calculating the frequency of discrete score across the evaluated users for each item, and outputting a predicted score distribution for each item;
Calculating discrete scores for all score data and outputting a predicted score distribution for each of the user and the item;
The collaborative filtering method according to claim 1, further comprising:   The collaborative filtering method according to claim 1, wherein the dissimilarity calculation step calculates the dissimilarity by calculating a KL divergence between the experience score distribution and the predicted score distribution.   2. The collaborative filtering method according to claim 1, wherein in the score estimation step, the optimization calculation is performed so that a total sum of the amount of the Cullback information between the experience score distribution and the predicted score distribution is minimized. A collaborative filtering device that predicts a score of an unevaluated item of a user using a score history including a score evaluated by a large number of users with a discrete value for a large number of items,
An experience score distribution calculating unit that outputs an experience score distribution indicating a distribution of evaluated scores based on the score history;
A predicted score distribution calculator that outputs a predicted score distribution by substituting a predetermined predicted initial value for an unrated score;
A dissimilarity calculator for outputting dissimilarities between the experience score distribution and the predicted score distribution;
A score estimator that performs non-linear optimization calculation so as to reduce the dissimilarity, and sequentially estimates and outputs unevaluated scores; and
A collaborative filtering device comprising:   A collaborative filtering program that causes a computer to execute the collaborative filtering method according to claim 1.   A recording medium in which the collaborative filtering program according to claim 7 is recorded.

Images