KR100447137B1 - A extraction and prediction method of latent user's preference - Google Patents

Abstract

The present invention provides a potential for improving prediction accuracy and reducing execution time through noise reduction and meaningful dimension reduction of input matrix in extracting and predicting potential user preferences using a staged evaluation matrix inputted by users. A method for extracting and predicting user preferences, the method comprising: extracting potential user preferences using a unique value analysis (SVD) technique, and performing noise removal and significant dimensional reduction of an input matrix; A second process of mapping users to a meaningful dimension space, but representing the users as a vector in the reduced dimension space; A third step of obtaining a distance representing similarity between each user by using a vector representing the users; A fourth step of filtering users with less similarity using a threshold of a predetermined size to select users similar to each user; A fifth step of obtaining weights to give more influence to more similar users; And a sixth process of obtaining the users' preferences for the items by using the average of the weighted evaluation values of all similar users.

Description

How to extract and predict potential user preferences {A EXTRACTION AND PREDICTION METHOD OF LATENT USER'S PREFERENCE}

The present invention relates to a method of extracting and predicting potential user preferences. More specifically, the present invention relates to a method of extracting and predicting user preferences for multimedia items using singular value decomposition (SVD) technology. The present invention relates to a method of extracting and predicting potential user preferences by using a stepwise evaluation matrix to extract and predict potential user preferences.

The present invention to be described below can be considered as an essential part of the technique of collaborative information filtration. Collaborative information filtration is a method of collaborative filtration based on a gradual evaluation of a user's items to find users with similar evaluation patterns and cross recommending the items that the users prefer.

In the conventional case, many methods are used for collaborative information filtration as described below.

The first method is to find the similarity among users using the correlation coefficient.

The method using the correlation coefficient uses a linear combination of the observed ratings for items that the user does not rate. The main idea of how to use correlation coefficients is to use formal similarity measures to give greater weight to users with similar preferences. Among the methods for obtaining the correlation coefficient, the representative method is the Pearson r method.

Pearson r in a pair of (x_i ~, ~ y_i), i = 1, ..., N is given by the following equation (1).

Secondly, vector similarity is used.

The similarity of two documents in information retrieval is measured by treating each document as a vector of word frequency and calculating the cosine angle formed by the two frequency vectors. Applying this form to collaborative filtration allows users to act as documents, items as words, and staged evaluation as word frequencies. The votes observed in this algorithm show positive preferences and negative votes play no role. Unobserved items have a value of zero.

Relevance weights are given by Equation 2 below.

Here, the squared term in the denominator serves as the normalization of voting so that users who vote for more items will not be deductively similar to other users.

Third, there is a clustering method.

One possible way is that the probability distribution of votes is conditionally independent given membership to class variable C that is not observed for a relatively small number of distributed values. The idea is that there are some groups or types of users who have a common set of preferences and tastes. Given a class, the preferences associated with the various items are independent. Probability distributions that relate the combined probability distribution of voting and class to a set of conditional and nearest distributions that are easy to handle can be expressed by Equation 3 below using the standard "naive" Bayes formula:

By the way, the authentication methods used in the prior art as described above has a problem that is difficult to apply in terms of execution time, prediction accuracy, potential semantic extraction.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and its purpose is to remove noise and semantics of input matrix in extracting and predicting potential user preferences by using a stepwise evaluation matrix inputted by users. The present invention provides a potential user preference extraction and prediction method that can increase prediction accuracy and reduce execution time through dimension reduction.

1 is a diagram illustrating a system for recommending a multimedia item using SVD based on the Internet according to the present invention.

FIG. 2 is a schematic block diagram illustrating the configuration of a system applied to FIG. 1.

3A-3C show database tables of a system according to the present invention.

4 is a description of the overall algorithm applied to the system of the present invention

5 is a flow diagram illustrating the overall operation of a system to perform a recommendation for a multimedia item.

According to a feature of the present invention for achieving the above object, a first process of extracting the potential user preferences using the unique value analysis (SVD) technology, the noise removal of the input matrix and the meaningful dimensional reduction ; A second process of mapping users to a meaningful dimension space, but representing the users as a vector in the reduced dimension space; A third step of obtaining a distance representing similarity between each user by using a vector representing the users; A fourth process of filtering out less similar users using a threshold of a predetermined size to select users similar to each user; a fifth process of obtaining weights to give more influence to more similar users; And a sixth process of obtaining the user's preferences for the items using the average of the weighted evaluation values of all similar users.

The above objects and various advantages of the present invention will become more apparent from the preferred embodiments of the invention described below with reference to the accompanying drawings by those skilled in the art.

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

1 is a view for explaining a system for recommending a multimedia item using SVD based on the Internet according to the present invention, FIG. 2 is a schematic block diagram for explaining the configuration of the system applied to FIG. 3C is a diagram showing a database table of a system according to the present invention, FIG. 4 is a description of an overall algorithm applied to the system of the present invention, and FIG. 5 is an overall operation of a system for performing a recommendation for a multimedia item. It is a flow chart showing.

Referring to FIG. 1, each user PC is connected to a web server through the Internet, and a DBMS and a preference extraction and prediction server are connected to the web server.

At this time, preference information for the item is obtained by receiving user information, item information, step-by-step evaluation information, and the like as shown in FIG. 2, wherein the user information has a form as shown in FIG. The information has a form as shown in FIG. 3B, the staged evaluation information has a form as shown in FIG. 3C, and the preference information has a form as shown in FIG. 3D. Similarly, user information, item information, and graded evaluation information are managed by a user information manager, an item information manager, and a graded evaluation information manager, respectively. A singular value decomposer receives user information, item information, and stepwise evaluation information from the user information manager, item manager, and rating manager, which will be described later. Perform unique value analysis (SVD) according to the method. The latent preference extractor extracts potential user preferences according to a method described below using the result of the unique value analysis. The preference predictor predicts item preferences of users according to a method described below. Recommendation Manager finds users who have similar evaluation patterns and manages preference information on their items to enable collaborative filtering.

4 and 5, the latent semantic structure has been used as a method for searching a document. Primarily, the matrix of terms and documents was analyzed by Singular Value Decomposition (SVD) to derive a particular potential semantic structure. The unique value analysis (SVD) is closely related to various fields of mathematics and statistics such as eigenvecter decomposition, spectral analysis, factor analysis, etc. As shown in FIG. 4, the overall algorithm applied to the system of the present invention is One. Unique value analysis of R, a stepwise evaluation matrix of users. Among the decomposed matrices, the matrix product of U and S is calculated to represent users as a vector in a reduced dimension space. A process of obtaining a cosine distance of all users based on k predetermined columns of US, that is, a distance representing similarity between users. Treat all users with co-distance greater than a certain threshold T as similar users, N _u , ie filter out less similar users using a threshold of a certain size to select similar users to each user. Process of 5. The process of calculating weights among similar users, that is, the process of obtaining weights to give greater influence to more similar users. Calculating the average of the weighted evaluation values of all the similar users, and calculating the user's preference for the items j.

The unique value analysis (SVD) starts with an arbitrary rectangular matrix with different presences in rows and columns. For example, a row can be a user and a column can be an item. This orthogonal matrix can be decomposed into three other matrices of a particular type by a process called unique value analysis (SVD). Linearly, most of the independent elements will be very small, or can be ignored, resulting in an approximation model with smaller dimensions. In this reduced model, the user-to-user, item-to-item, and user-item similarity is estimated to be reduced in value. As a result, the inner product, or cosine distance, of the two vectors is represented in a geographic spatial configuration corresponding to the measured similarity.

For collaborative filtration, SVD can be viewed as a technique for removing a collection of items that are irrelevant in distinguishing a user's characteristics. Each user is represented as a vector of evaluation values. As an effect of dimension reduction, users with somewhat different profiles can be located as the same vector. This is an attribute for achieving incredible improvements in data.

You can reconstruct the matrix of original users and items from the weights of the elements, ideally but not completely accurate. Since the original user space is not reliable, it should be noted that the derived k-dimensional element space does not completely reconstruct the original user space. It simply requires a derived structure that expresses reliability and importance by using items as pointers to users.

Conceptually, the unique value analysis (SVD), a special model of potential structure, shows that u * i, a rectangular matrix representing u users and i items, can be broken down into three different matrices.

R = UST ^ T

Where U and V are orthogonal matrices and S is a diagonal matrix. This form of decomposition is called unique value analysis (SVD) of the R matrix. U and V are called left and right singular vectors, respectively, and S is a diagonal matrix of singular values.

The following figure shows the schematic of the unique value analysis (SVD) for a matrix of users versus items. That is, it is a conceptual diagram of unique value analysis (SVD) of the evaluation matrix R. FIG.

[Model 1]

In general, SVD provides a simple method for optimal approximation using smaller matrices. If sorted by the magnitude of the singular value of S, the k largest values are retained and the remaining small values are set to zero. As a result, the product of the matrix is defined as R 'which is approximately equal to R and is represented by Equation 4.

Including the k largest independent linear elements of R has the effect that R 'produces an importantly relevant structure of the data and removes most of the noise. The following figure shows a reduced model that approximates the evaluation matrix.

[Model 2]

In selecting the reduced dimension, k, it should be chosen large enough to fit the actual structure of the data and small enough not to indicate errors in sampling and insignificant detail.

In general, decomposition of SVD allows comparison of the similarity between a user and a user, a user and an item, and an item and an item. However, in the collaborative filtration, the similarity between users is the most important factor, so we will focus on comparing the similarity between users.

The dot product of the two vectors of R 'reflects the degree to which the two users evaluate similar items for the item. R'R ' ^T, which contains the product of all users versus users, is a symmetric square matrix. Since S is a diagonal matrix and U is an orthogonal matrix, R'R ' ^T can be calculated by Equation 5 as follows.

Point in R'R ^'T (i, j) can be obtained by taking the dot product of the i and j row of the matrix US. That is, considering the column of TS as the coordinates for the user, the dot product of the two points means the comparison of the users.

The dot product of two points may be used as a criterion indicating the similarity of users, or the cosine distance of two vectors may be used. The matrix product of US constitutes a vector representing the similarity of the evaluation patterns of the two users. That is, the smaller the cosine distance of the two vectors, the greater the similarity of the two users.

While the invention has been shown and described in connection with specific embodiments thereof, it is well known in the art that various modifications and changes can be made without departing from the spirit and scope of the invention as indicated by the claims. Anyone who owns it can easily find out.

As described above, according to the potential user preference extraction and prediction method of the present invention, the noise extraction of the input matrix and the meaningfulness of the input matrix are significant in the extraction of the preferences by extracting and predicting the potential user preferences using the stepwise evaluation matrix inputted by the users. By reducing the dimension, it is possible to improve the prediction accuracy and reduce the execution time than using Pearson r or neural network.

Claims (1)

In a system consisting of a user PC, a web server connected to the user PC through the Internet, a DBMS connected to the web server, and a preference extraction and prediction server, the user can extract and predict potential user preferences by the preference extraction and prediction server. In the way, A first process of extracting potential user preferences using singular value decomposition (SVD) technology and performing noise removal and significant dimensional reduction of the input matrix; A second process of mapping users to a meaningful dimension space, but representing the users as a vector in the reduced dimension space; A third step of obtaining a distance representing similarity between each user by using a vector representing the users; A fourth step of filtering users with less similarity using a threshold of a predetermined size to select users similar to each user; A fifth step of obtaining weights to give more influence to more similar users; And a sixth step of obtaining the user's preferences for the items using the average of the weighted evaluation values of all similar users.