JP2013109425A - Content recommendation device and method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To personalize and recommend an item (content) in accordance with the interest of a user at that time, and in accordance with a trend.SOLUTION: In the content recommendation device, the number of consumption of contents by a user is separated at regular time intervals from a consumption history to which time is added, and expressed by vector. The sum of the vectors is calculated for each class of the contents to generate the vectors for each class and for each user. The latest vector of a recommended user and a part of the past vectors having the same time length of another user are obtained. The similarity between a pattern of the vector of the recommended user and a pattern of the vector of another user is obtained. A content to be recommended is determined by using the similarity of the patterns held by the users and the contents exist in the patterns.

Description

  The present invention relates to a content recommendation apparatus, method, and program, and in particular, considers a time series, detects a product (content) that a user is interested in, and a change in a class to which the content belongs, and recommends a product that matches that time. The present invention relates to a content recommendation device, method, and program.

  On the Web network, systematic dictionaries such as WikiPedia (registered trademark) have become widespread as requests for reference to the meaning and concept of objects by users have increased. Also, in order to enable these services to be processed by users instead of humans, and to be presented to users after customization, machine-processable concept reference APIs (Application Program Interface) are rapidly spreading. , DBPedia (Registered Trademark), Word-Net (Registered Trademark), FreeBasse (Registered Trademark) and other information providers have come to systematize their information and present them at low cost and free through API. (For example, see Non-Patent Document 1).

  On the other hand, a recommendation system for guessing a concept of interest of a user and collecting and presenting information on behalf of the user is also required and studied. Combining these studies with the various APIs above, you can estimate the user's interests in a wider range. However, there are two major problems with the current technology.

  (1) It seems that the change in interest according to the time axis is more severe as the variety of interests are handled, but it does not support such conversion.

  (2) The concept system is frequently maintained. For example, the video system is deepened with the deepening of services, and when a new group gains popularity, the system of the group can be created, and the system also deepens. It does not take into account such changes in the conceptual system (taxonomy).

  In the present invention, the problem (1) is addressed.

  The technique for estimating interest is shown below.

  First, interest estimation based on taxonomy will be described.

  It does not deal with changes in user interest, and even if it is a history two years ago, it is treated as equivalent to the recent history. In particular, when managing user interests with abstract information such as classes, if the time series becomes longer, it becomes easier to assume that users are interested in a very large number of classes (for example, non-patent literature). 2).

  Next, the entire consumption history is used instead of using the consumption history leaked from the window by measuring the user's interest in a specific time window. We have built a model that takes into account the characteristic that the average value of user ratings for items depends on the date and time, and is applied to two methods: Matrix Factorization and k-nearest users / items Therefore, the accuracy of both conventional MF and KN techniques is improved (see, for example, Non-Patent Document 3).

Linked Open Data Project (http://linkeddata.org/). Makoto Nakatsuji, Yasuhiro Fujiwara, Akimichi Tanaka, Tosho Uchiyama, Ko Fujimura, Toru Ishida, "Classical Music for Rock Fan", The 19th ACM international conference on Information and knowledge management. Collaborative Filtering Temporal Dynamics, by Yehuda Koren

  However, in the method of Non-Patent Document 2, as shown in FIG. 1, it is similar to other users in many classes, and the accuracy may decrease. On the other hand, as shown in FIG. 2, there is a problem that only the recent history is too sparse and the accuracy is deteriorated.

  Further, the method of Non-Patent Document 3 does not clearly handle the items of interest or the genre changes themselves. Furthermore, it is not suitable for purchase history because it is a method only for grades.

  The present invention has been made in view of the above points, a content recommendation device and method capable of personalizing and recommending items (contents) in line with the user's interest at the time and in accordance with the trend, and The purpose is to provide a program.

In order to solve the above problem, the present invention (Claim 1) is a content recommendation device that recommends content that a user is interested in at the time,
Consumption history storage means for storing a consumption history of a user's content having an evaluation value for the user's content and time information evaluated;
Time giving means for converting the time information of the consumption history into a time by giving the consumption history and the taxonomy of the content obtained from the information source of the consumption history storage means as input; and
User interest construction means for extracting the user's interest from the consumption history given the time, structuring the data and storing it in the user interest storage means;
Similar pattern extraction means for extracting a consumption trend pattern of other users similar to the recommended user from the data structured user interest of the user interest storage means, and storing it in the similarity storage means;
Prediction value measuring means for determining content recommended to the user based on the occurrence frequency of the content in the similarity pattern of the similarity storage means.

Further, the present invention (Claim 2) is the user interest construction means,
The number of times the user's content has been consumed is expressed as a delimited vector at regular time intervals from the consumption history given the time, and the vector sum for each class of content is determined to generate a vector for each class user. Means for storing in the means,
In the similar pattern extracting means,
The latest vector of the recommended user and a part of the past with the same time length of the other user are acquired from the user interest storage means, and the vector pattern of the recommended user and the vector pattern of the other user Means for obtaining a similarity and storing it in the similarity storage means;
In the predicted value measuring means,
Means for determining the recommended content using the similarity of the pattern held by the user of the similarity storage means and the content existing in the pattern;

  According to the present invention, it is possible to personalize and recommend items in line with interests at the time and along with trends by recommending user content whose latest history is similar to past history. It becomes. Since class information is also taken into account, it can also have a basis for recommendation.

It is a figure which shows the conventional trouble (The relationship between users cannot be measured only by item and sparse). It is a figure which shows the conventional problem (similar log is sparse only in t vicinity). It is a figure which shows the characteristic (searching for a similar pattern also in the past log) of this invention. It is a block diagram of the content recommendation apparatus in one embodiment of this invention. It is a flowchart of the process of the content recommendation apparatus in one embodiment of this invention.

  First, the technology that forms the basis of the present invention will be described.

In the traditional collaborative filtering, when measuring the similarity between users, Cosine-based approach (Breese, Heckerman, & Kadie 1998, Sarwar, Karypis, Konstan, & Ridel 2001) and Pearson correlation approach (Resnick, Iacovou, & Suchak 1994, Shardanand & Maes 1995) are often used. In the Cosine-based approach, the similarity S (a, u) between the active user a and a user u is obtained by calculating the Cosine angle of the evaluation vectors of a and u. Formally, let M be the number of items that both users have given an evaluation, and let the evaluation value for item I _i of user u be

とすると、類似度Ｓ（ａ，ｕ）は以下の式(１)で与えられる。

Then, the similarity S (a, u) is given by the following equation (1).

一方、Pearson correlationアプローチでは、ユーザ間の類似度の計測において、ユーザの評価スキームはユーザ毎に異なるという考え方を採用しており、以下の式(2)で表される。

On the other hand, the Pearson correlation approach adopts the concept that the user evaluation scheme differs for each user in measuring the similarity between users, and is expressed by the following equation (2).

ここで、

here,

は、ユーザｕのアイテムへの評価値の平均値を示す。

Indicates an average value of evaluation values for the items of the user u.

  It should be noted that both the Cosine-based approach and the Pearson correlation approach focus on and calculate only items that are rated by both users a and u.

If N is the number of users having high similarity to the active user a, the predicted value for the item I _{i of} a

は、以下の式（3）で与えられる。

Is given by the following equation (3).

以下図面と共に、本発明の実施の形態を説明する。

Embodiments of the present invention will be described below with reference to the drawings.

  In the present invention, the following two points are the gist.

  (1) It is based on the idea that the user's interest fluctuations that are too sparse at the item level can be seen abstractly at the class level. This is because the class aggregates items, so sparse item information can be closely observed in the class. For example, there are items i1, i2, and i3 under rock, and even if they are purchased once each, they can be considered to have been purchased three times in class. The change of the recent consumption of the recommended user (a) is grasped at the class (concept) level, and the users whose trends in the recent consumption change are similar are calculated. And a similar user recommends an item recently purchased.

  (2) If the recent consumption behavior is similar only from the current time, the log becomes too sparse and the accuracy is reduced (FIG. 2). That is, since the log becomes too sparse and the accuracy deteriorates only with the recent log whose recent consumption fluctuation is close to the user a, in the present invention, as shown in FIG. A past log pattern of a user close to is used when calculating a recommendation, and a recommended item for the current user a is calculated with high accuracy.

  FIG. 1 shows the configuration of a content recommendation device according to an embodiment of the present invention.

  The content recommendation device shown in FIG. 1 includes a consumption history DB 1, a user interest storage unit 2, a similarity storage unit 3, a time information addition unit 10, a user interest construction unit 20, a similar pattern extraction unit 30, and a predicted value calculation unit 40. Is done.

  The consumption history DB 1, the user interest storage unit 2, and the similarity storage unit 3 are storage media such as a hard disk.

  The consumption history DB 1 stores a user's consumption history (log) having time information evaluated as an evaluation value for the user's content. The user interest storage unit 2 is a user interest construction unit 20 and a similarity storage unit 3. Stores the data calculated by the similar pattern extraction unit 30.

  The time information adding unit 10 receives the information of the consumption history DB1 and the taxonomy that is the class hierarchy structure of the content obtained from the information source such as WikiPedia (registered trademark), and converts the time information of the consumption history of the content into time, This is given to the history information and output to the user interest construction unit 20.

  The user interest construction unit 20 constructs the user's interest (history information) as a data structure and stores it in the user interest storage unit 2.

  The similar pattern extraction unit 30 receives the user's interest structured as a data, extracts similar patterns of consumption trends of other users, and stores them in the similarity storage unit 3.

  The predicted value calculation unit 40 recommends the content to be presented to the user based on the content occurrence frequency among the similar patterns in the similarity storage unit 3.

  The operation of the content recommendation device having the above configuration will be described below.

  FIG. 2 is a flowchart of processing of the content recommendation device according to the embodiment of the present invention.

  Step 1) The time information assigning unit 10 reads history information from the consumption history DB 1 and assigns time information based on the taxonomy. In the present invention, the time when a user consumes an item (multimedia content such as music, a Web page, etc.) is separated at regular intervals and treated as discrete times. Specifically, with respect to the time information described in the history information, assuming that the first time on the history information is S and the value representing the period is D, each time t (i) is S + D × i To S + D × (i + 1).

  Step 2) The user interest building unit 20 data-structures the history information to which the time information is given in Step 1.

  1) A consumed item at a certain time t by the user u indicates an item consumed by the user u between S + D × t and S + D × (t + 1). If the item is consumed during the period, the order is not taken into consideration and the item is managed as a set of items consumed by the user u at time t.

2) A vector v _{u, i} is formed for each item i consumed by the user u. The column of the vector indicates time, and the element in the column stores the consumption frequency of the item i at each time.

3) A vector v _{u, c} is formed for each class c to which the item consumed by the user u belongs. The vector column indicates the time, and the element in the column stores the sum of the item consumption frequencies under c at each time. If v _{u, c} (t) is greater than 0, the value is propagated to the upper class of c. In order to reduce the amount of calculation, the value is not propagated if it is 0 or less. In particular, regarding items, if only zero or more items are managed as a set in advance, if only the set is checked and propagation is determined, the amount of calculation is reduced.

The value of the element corresponding to the vector v _{u, c} (t) is calculated as in equation (4). In equation (4), f (c) is a function that returns a set of child classes of c, and c ^s is a child class with c.

上記のように求めたｖ_u,c (t)をユーザ興味記憶部２に格納する。

The v _{u, c} (t) obtained as described above is stored in the user interest storage unit 2.

Step 3) The similar pattern extraction unit 30 reads v _{u, c} (t) from the user interest storage unit 2 and extracts a pattern similar to the recent history of the user a (recommended user).

1) The similar pattern extraction unit 30 particularly considers the consumption history up to the latest x times when viewed from the current time T for the recommended user a. Specifically, by setting only history information from T−x + 1 to T as a processing target, it is possible to realize a recommendation that emphasizes the consumption history at the latest x time. Therefore, the vector v _{u, c} (t) is degenerated to the latest time dimension. A consumption vector ra _{, c} for the class c in the consumption history at the latest x time is given. r _{a, c} is calculated as in the following equation (5). In equation (5), 0 indicates a zero matrix, and the subscript is the number of dimensions. E refers to the identity matrix, and the subscript is the number of dimensions.

２） 類似パターン抽出部３０は、被推薦ユーザａに対し、他ユーザの消費履歴で、クラスｃのベクトルr_a,cと類似するパターンをユーザ興味記憶部２から抽出する。当該処理は、アイテムレベルでのパターン抽出はスパースになり、不正確になるため、クラスレベルで実施する。また、計算量を削減するための工夫として、同じクラスを持たないパターンに関しては、その配下クラス以降のパターン抽出の計算をスキップする。

2) The similar pattern extraction unit 30 extracts a pattern similar to the vector r _{a, c} of the class c from the user interest storage unit 2 in the consumption history of other users for the recommended user a. This processing is performed at the class level because pattern extraction at the item level becomes sparse and inaccurate. As a device for reducing the amount of calculation, for patterns that do not have the same class, calculation of pattern extraction after the subordinate class is skipped.

First, the similar pattern extraction unit 30 calculates a pattern vector p _{u, t, c} from t−x + 1 to t as shown in Expression (6) with a time t of a certain user u as a final time.

３） 次に、類似パターン抽出部３０は、あるクラスｃに対するユーザａとあるユーザｕのある時刻ｔを最終時刻とするパターンp_u,tとの類似度を求める。当該類似度の計算式を式(7)に示す。式(7)において、cos()はcosine類似度を示す。Ｃはユーザｕに対し、上記の２）で計算済みのクラス集合を指し、タクソノミ上の全クラスではない（計算量削減のため）。

3) Next, the similar pattern extraction unit 30 obtains the degree of similarity between the pattern p _{u, t} whose final time is a certain time t between the user a and the user u for a certain class c. The formula for calculating the similarity is shown in Formula (7). In Equation (7), cos () indicates the cosine similarity. C indicates to the user u the class set calculated in 2) above, and not all classes on the taxonomy (to reduce the amount of calculation).

なお、cosine類似度ではなく、時系列の類似性を測る手法であれば、Ｎグラム法などを利用するのでもよい。

Note that the N-gram method or the like may be used as long as it is a method for measuring time series similarity instead of cosine similarity.

Note that the calculation of the similarity between the consumed items of the user a and the user u does not consider the time series in order to avoid the log becoming too sparse and to reduce the calculation amount. The similarity of the consumed items is calculated from the following equation (8) as in Non-Patent Document 2, where I _u is an item set at all times under the class c of the user u. In equation (8), jac () represents a jaccard coefficient.

なお、アイテムの類似度の計算方式は式(8)以外の公知の方法を用いてもよい。

The item similarity calculation method may be a known method other than Equation (8).

4) The similar pattern extraction unit 30 calculates the similarity S _{a, u, t} between the patterns of the user “a” and the user “u” according to the equation (9). In equation (9), N is a normalization function.

上記の式(9)により求められたパターン間の類似度を類似度記憶部３に格納する。

The similarity between patterns obtained by the above equation (9) is stored in the similarity storage unit 3.

  Step 4) The predicted value calculation unit 40 reads the similarity from the similarity storage unit 3 and extracts a content having a high occurrence frequency in the similar pattern. Specifically, the predicted value of the item to the user is calculated using the above equation (3). At this time, in Formula (3), each pattern similarity degree which a user has is substituted instead of a user similarity degree, and item Ii is an item which exists in the pattern which counters.

  The predicted value calculation unit 40 outputs the calculated predicted value of the item as a recommended item set.

  The differences between the present invention and Non-Patent Document 2 are shown below.

  (1) In the present invention, not all of the consumption histories possessed by the user are used for classes, but only a part of organizations in time series are used. For the recommended user (user a), only the latest consumption history is used, and for the similarity calculation target user (user u), only an arbitrary partial period in time series is used. As a result, it is possible to cope with fluctuations in user interests and avoid sparse problems, while avoiding the conventional problem that users who take various consumption behaviors over a long period of time are connected to a large number of users at the class level. it can.

  (2) In the present invention, as a set of items for predicting an item, the similarity for each user is not used, but only a portion with a similar consumption pattern of the user is used. As a result, items that deviate from the current user's interest can be excluded from the recommendation, and the relevance rate is increased.

  Next, the difference between the present invention and Non-Patent Document 3 will be shown.

  Non-Patent Document 3 does not introduce the difference in the user's interests into the prediction model, but introduces it into the model of the change in the average value of the user's rating and the change in the famousness of the item itself. , It does not consider the similarity of the variation of interest for each individual. Therefore, the recommendation is not presented with an emphasis on the current user's interest. The present invention can extract a past variation pattern corresponding to the current variation in interest for each individual, and determines a recommendation from the extracted variation pattern. Can recommend items.

  In the above-described embodiment, the content is described as an object. However, the present invention is not limited to this example, and can be applied to general products.

  In the present invention, each operation of the components of the content recommendation device in FIG. 1 described above can be constructed as a program, installed on a computer used as the content recommendation device, executed, or distributed via a network. It is.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

1 Consumption history database (DB)
2 User interest storage unit 3 Similarity storage unit 10 Time information giving unit 20 User interest construction unit 30 Similar pattern extraction unit 40 Predicted value calculation unit

Claims (5)

A content recommendation device that recommends content that the user is interested in at the time,
Consumption history storage means for storing a consumption history of a user's content having an evaluation value for the user's content and time information evaluated;
Time giving means for converting the time information of the consumption history into a time by giving the consumption history and the taxonomy of the content obtained from the information source of the consumption history storage means as input; and
User interest construction means for extracting the user's interest from the consumption history given the time, structuring the data and storing it in the user interest storage means;
Similar pattern extraction means for extracting a consumption trend pattern of other users similar to the recommended user from the data structured user interest of the user interest storage means, and storing it in the similarity storage means;
Predicted value measuring means for determining content to be recommended to the user based on the occurrence frequency of the content within the similarity pattern of the similarity storage means;
A content recommendation device comprising: The user interest building means is
The number of times the user's content has been consumed is expressed as a delimited vector at regular time intervals from the consumption history given the time, and the vector sum for each class of content is determined to generate a vector for each class user. Means for storing in the means,
The similar pattern extraction means includes:
The latest vector of the recommended user and a part of the past in which the time length of the other user is the same are acquired from the user interest storage means, and the vector pattern of the recommended user and the vector pattern of the other user Including a means for obtaining the similarity and storing the similarity in the similarity storage means,
The predicted value measuring means includes
The content recommendation apparatus according to claim 1, further comprising means for determining the recommended content using the similarity of the pattern held by the user of the similarity storage means and the content existing in the pattern. A content recommendation method for recommending content that a user is interested in at the time,
The time giving means inputs the consumption history of the consumption history storage means storing the evaluation value for the user content and the consumption history of the user content having the evaluated time information and the content taxonomy obtained from the information source as input. A time giving step for converting time information of the history into time and giving it to the consumption history;
A user interest building means for extracting the user's interest from the consumption history given the time, structuring the data and storing it in the user interest storage means; and
A similar pattern extracting step in which a similar pattern extracting unit extracts a consumption trend pattern of another user similar to the recommended user from the data structured user interest of the user interest storing unit and stores it in the similarity storing unit When,
A predicted value measuring step, wherein the predicted value measuring means determines content to be recommended to the user based on the occurrence frequency of the content in the similarity pattern of the similarity storage means;
The content recommendation method characterized by performing. In the user interest building step,
The number of times the user's content has been consumed is expressed as a delimited vector at regular time intervals from the consumption history given the time, and the vector sum for each class of content is determined to generate a vector for each class user. Stored in the means,
In the similar pattern extraction step,
The latest vector of the recommended user and a part of the past with the same time length of the other user are acquired from the user interest storage means, and the vector pattern of the recommended user and the vector pattern of the other user Find the similarity, store it in the similarity storage means,
In the predicted value measurement step,
The content recommendation method according to claim 3, wherein the recommended content is determined by using the similarity of the pattern held by the user of the similarity storage means and the content existing in the pattern. Computer
The content recommendation program for functioning as each means of the content recommendation apparatus of Claim 1 or 2.

Images