KR101442122B1 - Apparatus and Method for Content Recommendation - Google Patents

Abstract

The present invention discloses a content recommendation apparatus and method. A content recommendation apparatus according to an aspect of the present invention includes a collection unit for collecting rating data corresponding to a first type of content to be recommended by a user in online content rating data having a power law distribution, A preprocessing unit for selecting rating data of users rated for the first type of content and having a number of evaluation times equal to or greater than a threshold number of ratings, a rating unit for predicting non-predefined ratings using the selected rating data, A rating unit for calculating overall rating data reflecting the ratings and predicted ratings of all the screening target users for all the content items to be recommended corresponding to the recommendation target contents, Provides a predefined number of users from the highest rated content in the content It is characterized in that it comprises like parts.

Description

[0001] APPARATUS AND METHOD FOR CONTENT RECOMMENDATION [0002]

The present invention relates to content recommendation, and more particularly, to a content recommendation apparatus and method capable of recommending content using user's rating data.

Recently, Word of Mouth in online social networks such as Epinons, Facebook, and IMDB has had a great impact on people's purchasing and movie viewing.

In addition, online movie ratings, which are part of the film's marketing strategy, are attracting the attention of those who are looking for high-quality movies.

For example, ratings of movie rating sites such as Movielens, Rotten Tomato, Daum and Naver in foreign countries are influencing viewers' choice of movies.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a content recommendation apparatus and method for predicting a rating of a content using rating data having a distribution according to a power law.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood by those skilled in the art from the following description.

A content recommendation apparatus according to an embodiment of the present invention includes: a collection unit for collecting rating data corresponding to a first type of content to be recommended by a user in online content rating data having a power law distribution; A preprocessing unit for selecting rating data of users rated for the first kind of content and having a number of evaluations equal to or greater than a threshold number of the collected rating data; Estimating a score not pre-determined using the selected rating data, and calculating overall rating data reflecting the ratings and predicted ratings of all the selected users for all the recommended content corresponding to the first type of content A prediction unit; And a recommendation unit for averaging the ratings of the content to be recommended in the calculated overall rating data and providing a predetermined number of users from the highest rated content among all the recommended content.

According to another aspect of the present invention, there is provided a method of recommending content by an apparatus, comprising: collecting rating data corresponding to a first type of content to be recommended by a user in content rating data on-line having a power law distribution; Selecting rating data of the users rated for the first type of content and the rating number of which is equal to or greater than the threshold number of ratings collected from the collected rating data; Estimating a score not pre-determined using the selected rating data, and calculating overall rating data reflecting the ratings and predicted ratings of all the selected users for all the recommended content corresponding to the first type of content step; And averaging the ratings of the content to be recommended from the calculated total rating data and providing the user with a predetermined number of contents from the highest rated content among all the recommended content.

According to the present invention, it is possible to improve the accuracy of prediction of a score using rating data.

FIG. 1 and FIG. 2 are graphs showing In-degree distribution and Out-degree distribution of Daum movie rating network.
3 is a view for explaining a matrix factorization method;
4 is a configuration diagram showing a content recommendation apparatus according to an embodiment of the present invention;
FIG. 5 is a graph illustrating a comparison of the accuracy of the rating prediction according to the embodiment of the present invention with other rating prediction methods when the threshold number is set to 5, and FIG.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. As used herein, the terms " comprises, " and / or "comprising" refer to the presence or absence of one or more other components, steps, operations, and / Or additions.

Before describing the specific configuration of the present invention, the movie rating network will be analyzed.

To this end, we crawled ratings data of Daum movie rating network from February 25, 2009 to May 14, 2012. Here, crawling is a technique of collecting documents stored in a large number of computers and storing them as an index of a search target.

In the Daum movie rating data of the period (20009-02-25 ~ 2012-05-14), the total number of movies is 11682, the number of users is 140271, and the number of the users in the movie is 425537.

1 and 2 are graphs showing in-degree distribution and out-degree distribution of Daum movie rating network. Herein, the rating is the number of users who evaluated one movie, and the out rating is the number of movies evaluated by one user.

1 and 2, the Y axis is a complementary cumulative density function (CCDF), which is equal to 1 minus the cumulative density function (CDF) value.

For example, in FIG. 1, when x = 1000, y = 0.01, which means that the proportion of movies that received more than 1000 evaluations is 0.01 in total.

In Fig. 1 and Fig. 2, the fitting line (see the dotted line) is a graph according to the power law, and its distribution follows Pr (X = x) = ax, and the b value is the absolute value of the slope.

In FIGS. 1 and 2, it can be seen that the Daum movie rating network follows a power function distribution such as a degree distribution of neighbors having a relationship with a user in an online social network.

Through this, it can be seen that online movie rating network has characteristics of online social network.

Hereinafter, a method of recommending content based on an online content rating network will be described. At this time, the content includes a movie, a drama, a movie, a novel, a photo album, and the like. In the following description, a method for recommending a movie based on a movie rating network will be described for convenience of explanation.

<< Matrix factorization method >>

The movie can be recommended by Matrix Factorization which applies SVD (Singular Value Decomposition) method to the rating data of online movies.

In the matrix factorization method, the predictive rating data of the movie i of the user u can be calculated as shown in Equation (1) below.

In Equation 1, p _u is a feature of the entire user, and q _i is a feature of the entire movie, and each has a relationship as shown in FIG. As shown in FIG. 3, R _{m × n} is total rating data, each user u is in the relationship of vector p _u ∈ R ^f , and each item (movie) i is in the relationship of vector q _i ∈ R ^f . Where d _i is a latent factor. As described above, by using the matrix factorization method, it is possible to predict the unrated scores (see? In Fig. 3) from the relationship between the user u and the item i.

For more detailed matrix factoring methods, see S. Deerwester, S. Dumais, GW Fumas, T. K Landauer and R. Harshman, "Indexing by Latent Semantic Analysis", Journal of the Society for Information Science 41 1990), pp. 391-407), and a description thereof will be omitted.

The matrix factorization method provides relatively good performance, but if mixed with other methods, the performance can be further improved.

<< Average bias  How to >>

The average bias method considers the bias b _ui that reflects the characteristics of the user and the item, as shown in Equation 2 below.

In equation (2), u is the average rating of the entire movie, b _i is the difference between the average rating of movie i and u, and b _u is the difference between the average rating of user u and u.

For example, if the average rating of the entire movie is 5, the average rating of the movie "Titanic" is 8, and the average rating of the withdrawal is 4, then b _i is 3 and b _u is -1. Therefore, b _ui becomes 7 (5 + 3-1) points.

The matrix factorization method and the average bias method can be used together to improve the performance of the content recommendation, which is described in Y. Koren, R. Bell, C. Bolinsky, "Matrix Factorization Techniques for Recommender systems" Computer 2009).

<< Random bias  >>

The random bias method according to the present invention is a method of reflecting the characteristics of the distribution of ratings of users and movies as shown in Equation (3).

In Equation (3), u is the average rating of the entire movie, r _i is the difference between u and u, a random variable having a distribution of ratings received by movie i, r _u is a random variable having a distribution of ratings of u, to be.

For example, if the average rating of an entire movie is 5, and the rating received by Titanic is 50% for 5 points and 50% for 10 points, create a random variable with 5 and 10 probability of 50%, subtract u from it, _i . r _u can be obtained in the same way.

A user with the same average score through the above-described method, or a movie, can have different results depending on its distribution. Typically, the user's rating tends to be biased, and in such cases correction of the shape according to the random bias may be effective.

According to the present invention, the performance of content recommendation can be improved by using a matrix factorization method and a random bias method.

As described above, in the online movie rating network of Figs. 1 and 2, the total data of a specific period is 140271 evaluators, 11682 movies, and 140271 ratings for movies. However, a sparse data set can be used when one or two users of the entire data occupy most of the users. If you use the matrix factorization method when there are a large number of users with low ratings, a cold start problem may occur.

In order to prevent such a problem, in the embodiment of the present invention, a movie is recommended on the basis of rating data of a user who rated a movie having a threshold value or more.

Also, if a user scored a rating of 0, the movie was strongly deprecated, and a score of 0 would not be rated by the user and would not be reflected in the user item matrix. In this case, a large error may occur in the evaluation result of the rating.

In order to prevent such a problem, in the embodiment of the present invention, one point is added to all the rated points of each user so that the rating is within the range of 1 to 11, and a value obtained by subtracting 1 from the predicted rating is used for movie recommendation.

In the example described above, the content is a movie, but it is also possible to recommend the content after predicting the rating through other processes such as drama and novel.

Hereinafter, a content recommendation apparatus according to an embodiment of the present invention will be described with reference to FIG. 4 is a configuration diagram showing a content recommendation apparatus according to an embodiment of the present invention.

4, the content recommendation apparatus 10 according to an exemplary embodiment of the present invention includes an input unit 110, a collecting unit 120, a preprocessing unit 130, a predicting unit 140, and a recommending unit 150, . At this time, it goes without saying that the input unit 110 may not be provided in the content recommendation apparatus 10, but may be provided outside the input unit 110.

The input unit 110 is a user interface, and selects a type of content to be recommended by the user. For example, the types of contents are movies, dramas, novels, cartoons, and the like. At this time, the user can select to receive a movie through the input unit 110.

The collection unit 120 collects content rating data of a kind selected by a user from content rating data on an online network having a power function distribution.

Here, the content rating data of the online network having the power function distribution may be content rating data of a site providing movie ratings such as Daum, Naver, and a movie information providing site.

The preprocessing unit 130 selects content rating data of a user who rated a homogeneous content equal to or more than a threshold number among the collected content rating data. Here, the critical number is preferably 6 or more.

For example, the preprocessing unit 130 selects content rating data of a screening user who has rated a movie having a threshold number or more from the movie rating data.

The preprocessing unit 130 adds 1 to the selected content rating data to make the rating scale from 0 to 10 to 1 to 11. [ In the embodiment of the present invention, in this way, it is possible to distinguish the rated points from the user and the unrated content.

The predictor 140 predicts the content rating data that is not rated from the scaled content rating data by matrix factorization and the random bias method as shown in Equation (4) below, and subtracts 1 from the predicted rating And calculates a predictive rating.

Here, a is the weight of the matrix factorization methods and the random bias (0≤a≤1), R _MF may be calculated by the equation (1) as a result of applying the matrix factorization to the content rating data, R _bias is the content Can be calculated by Equation (3) as a result of applying the random bias method to the rating data. In this case, the weights are preferably set through comparison of the matrix factorization and the average bias method. This will be described later with reference to Figs. 5 and 6. Fig.

For example, if the content rating data includes the ratings of the user 1 for the movies A, B, and C and the ratings for the movies A and C of the user 2, the predictor 140 uses the content rating data Predict the rating that User 2 expects to put on B, and subtract 1 from it. Accordingly, the prediction unit 140 can calculate the overall rating data in which all the rated users or the predicted ratings of all the target users are ranked with respect to all the recommended targets (for example, during or during the sale) corresponding to the selected type.

The recommendation unit 150 can recommend the content to the user who has requested the recommendation of the selected type of content using the comprehensive rating data.

For example, the recommendation unit 150 may average ratings of the movies in the comprehensive rating data for a plurality of movies, and recommend three movies to the user in the order of the highest rating.

Hereinafter, with reference to FIG. 5 and FIG. 6, the accuracy of the evaluation of the score according to the weight according to the embodiment of the present invention will be described with reference to RMSE (Root Mean Square Error) as a performance evaluation index.

FIG. 5 is a graph illustrating the accuracy of the score prediction method (MF Random) according to an embodiment of the present invention when the threshold number is set to 5, and FIG. 6 is a graph showing the number of thresholds FIG. 5 is a graph illustrating the accuracy of the prediction of a rating according to an embodiment of the present invention, in comparison with other rating methods. Here, the other rating prediction method is a method of combining matrix factorization (MF) and matrix factorization and average bias (MF Average).

In Figures 5 and 6, 90% of the total rating data is set to the training set (i.e., input data) and the rating is predicted for 10% that is not set to the training set by the training set. Then, the accuracy of rating prediction was verified by calculating the RMSE value between the predicted rating and the 10% rating data not set in the training set.

Here, for the convenience of evaluation, the entire rating data is data in which all target users are rated for all the recommended contents. In this case, the smaller the RMSE value, the higher the recommendation accuracy will be.

In FIG. 5, when the weight a is 0.5, the RMSE value of the MF average method is 2.385, which is the minimum value (Min), which is about 10.2% lower than the RMSE value 2.5 of the matrix factorization method.

In addition, the RMSE value of the MF average method of the present invention is 2.385 when the value of a is 0.1, which is 4.8% higher than that of the conventional matrix factorization method.

However, when the minimum value of the RMSE is compared, it can be seen that the MF Random method according to the embodiment of the present invention is less accurate than the average bias combining method (MF Average) when the threshold number is 5. However, as shown in FIG. 6, when the threshold number is greater than or equal to a certain degree, the accuracy of the rating prediction of the present invention becomes high.

6, the RMSE value of the MF Random method is 2.276 when the weight a is 0.1, which is smaller than the minimum RMSE value of the MF average method of 2.287 and the minimum RMSE value of 2.372 of the matrix factorization method of 0.4% 4.1% Accuracy is higher.

From this, it can be seen that the MF Random method according to the embodiment of the present invention further improves performance when the number of user ratings is large. Therefore, in the present invention, it is preferable to apply a threshold number of 6 or more.

As described above, the embodiment of the present invention can improve the accuracy of the prediction when the score is predicted using the rating data following the power function distribution.

While the present invention has been described in detail with reference to the accompanying drawings, it is to be understood that the invention is not limited to the above-described embodiments. Those skilled in the art will appreciate that various modifications, Of course, this is possible. Accordingly, the scope of protection of the present invention should not be limited to the above-described embodiments, but should be determined by the description of the following claims.

Claims (7)

A collection unit for collecting rating data corresponding to a content of a first type that a user desires to receive from content rating data on-line having a power law distribution;
A preprocessing unit for selecting rating data of users rated for the first kind of content and having a number of evaluations equal to or greater than a threshold number of the collected rating data;
Estimating a score not pre-determined using the selected rating data, and calculating overall rating data reflecting the ratings and predicted ratings of all the selected users for all the recommended content corresponding to the first type of content A prediction unit; And
A recommendation unit that averages the ratings of the recommendation target content from the calculated overall rating data and provides the users with a predetermined number of the content with the highest average rating among all the recommendation target contents,
And the content recommendation apparatus. The method according to claim 1,
The preprocessing unit adjusts the scale of the collected rating data by adding 1 to all of the collected rating data,
Wherein the predicting unit calculates the final predictive rating by subtracting 1 from the predicted rating and calculates the overall rating data by reflecting the calculated final predictive rating
/ RTI &gt; delete 2. The method of claim 1,
A content recommendation apparatus as claimed in any one of the preceding claims, Collecting rating data corresponding to a first type of content to be recommended by a user from online content rating data having a power law distribution;
Selecting rating data of the users rated for the first type of content and the rating number of which is equal to or greater than the threshold number of ratings collected from the collected rating data;
Estimating a score not pre-determined using the selected rating data, and calculating overall rating data reflecting the ratings and predicted ratings of all the selected users for all the recommended content corresponding to the first type of content step; And
Averaging the ratings of the recommendation target content from the calculated overall rating data and providing the user with a predetermined number of the content with the highest average rating among all the recommendation target contents
And the content recommendation method. 6. The method of claim 5,
The step of selecting the rating data includes:
Adjusting the scale of the collected rating data by adding 1 to all of the collected rating data,
Wherein the calculating step comprises:
Calculating a final predicted score by subtracting 1 from the predicted score, and calculating the overall score data by reflecting the calculated final predicted score
. delete

Images