JP2013109427A - Content recommendation device and method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recommend an item in accordance with the interest of a user at that time and a trend.SOLUTION: In the content recommendation device, time information of a consumption history is converted into time to be added to the consumption history, with taxonomy of a content obtained from the consumption history and an information source as an input. The consumption history to which the time is added is referred to. The consumption number of items consumed by the user is separated at regular time intervals to be expressed by vector. The sum of the vectors is calculated for each class to which the items belong, to generate the vectors for each class and for each user. The similarity between the users on the number of consumption of items between two times closest to the present time is calculated from the vectors for each class and for each user. The transition probability of consumption between the times is calculated on the basis of patterns of consumption of items at time (t-1) and at time t. An item to be recommended is determined by applying the transition probability to time (t+1).

Description

  The present invention relates to a content recommendation apparatus, method, and program, and in particular, in order to recommend content based on collaborative filtering, captures the content that the user is interested in and changes in the class to which the content belongs. The present invention relates to a content recommendation apparatus, method, and program for recommending and presenting content.

  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. In addition, machine-processable concept reference APIs (Application Program Interfaces) are rapidly spreading because services such as user requests are processed by humans instead of humans and can be customized and presented to users. , 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) The user's interest changes dynamically. The more diverse the interests are handled, the more the interest changes according to the time axis.

  (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.

  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, a method based on tensor decomposition will be described.

  In the method based on the Markov model for predicting the item purchased at (t + 1) from the relationship between the item purchased at the current time (t) and the item purchased at the previous time (t−1), Only the statistics based on purchases would predict (t + 1) items and are not personalized. Therefore, personal items can be personalized by purchasing similar items in (t-1), clustering user groups that have purchased similar items in (t), and recommending items purchased by such users. The item recommendation according to the Markov model is enabled (for example, see 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 Fans ?: Novel Recommendations for Expanding User Interests", pp. 949-958. . Factorizing Personalized Markov Chains for Next-Basket Recommendation by Stefffen Rendle (http://www.ra.ethz.ch/CDstore/www2010/www/p811.pdf)

  However, the above-described interest estimation based on the taxonomy is similar to other users in many classes, and accuracy may be reduced. On the other hand, the recent history alone has a problem that the history becomes too sparse and the accuracy is deteriorated.

  Also, in the method based on tensor decomposition, the fact that the same item was purchased is based on the cluster, so there is a problem that it takes time and effort for parameter tuning in clustering, and the generated cluster is a system for the user. It has a problem that it cannot be referred to.

  The present invention has been made in view of the above points, and an object of the present invention is to provide a content recommendation apparatus, method, and program capable of recommending an item that matches a user's current interest and trend.

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;
For the user interest structured in the data of the user interest storage means, measure the consumption pattern of similar users between two adjacent times, and store the similarity in the consumption pattern group storage means,
In the consumption pattern group similar to the consumption pattern group storage means, the transition probability calculation means for calculating the transition probability for each individual of the content or class and storing it in the transition probability storage means;
Prediction value calculation means for acquiring each transition probability in the transition probability storage means and calculating a recommended value of content for an individual.

Further, the present invention (Claim 2) is the user interest construction means,
Refer to the consumption history to which the time is given, and represent the number of items consumed by the user as a vector by dividing the number of consumptions at regular time intervals. Means for generating and storing in said user interest storage means;
In the similarity measuring means,
From the user-specific vector for each class of the user interest storage means, obtain the similarity between users of the number of times the item is consumed between the two most recent times of the current time, and the consumption pattern group storage means as consumption patterns of similar users Including means for storing in
In the transition probability calculation means,
Means for obtaining a transition probability of consumption between times based on consumption patterns of items at time (t−1) and time t from the consumption pattern group storage means;
In the predicted value calculation means,
Means for determining a recommended item by applying the transition probability at time (t + 1);

  According to the present invention, it is possible to personalize and recommend items (contents) in accordance with the interest at that time and in line with the trend. Also, it can have a recommendation basis by considering the class information.

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 calculation based on the class is introduced instead of the similar calculation based on the cluster while following the concept of the method based on the tensor decomposition.

  In the method based on the conventional Markov model that predicts the item purchased at (t + 1) from the relationship between the item purchased at the current time (t) and the item purchased at the previous time (t−1), The item of (t + 1) is predicted only from the statistics based on the purchase of the user, and is not personalized.

  Therefore, in the present invention, the same item / class is purchased in (t-1), the similarity with the user group who has purchased the same item / class in (t) is calculated, and among the users, Based on the item / class purchased at (t-1) and (t), the consumption probability of the item / class at (t) when the item / class at (t-1) is consumed is calculated.

  That is, the conditional probability of purchasing an item / class at (t) under the condition that the item / class at (t−1) is purchased is calculated. Based on the conditional probability, an item that can be purchased by the user a of the recommended person is calculated at (t + 1).

  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 time information adding unit 10, a user interest construction unit 20, a similarity calculation unit 30, a transition probability calculation unit 40, a predicted value calculation unit 50, a consumption history DB1, a user interest storage unit 2, and a consumption pattern. It comprises a group storage unit 3 and a transition probability storage unit 4.

  The consumption history DB 1, the user interest storage unit 2, the consumption pattern group storage unit 3, and the transition probability storage unit 4 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, and includes a user interest storage unit 2, a consumption pattern group storage unit 3, and a transition probability storage unit 4. Stores the data calculated by the user interest construction unit 20, the similarity calculation unit 30, and the transition probability calculation unit 40, respectively.

  The time information adding unit 10 reads a consumption log from the consumption history DB 1, acquires a taxonomy of content obtained from an information source such as WikiPedia (registered trademark), converts time information of the consumption history into time, and converts the time into the consumption history. And output to the user interest construction unit 20.

  The user interest constructing unit 20 constructs the user's interest from the consumption history given the time acquired from the time information giving unit 10 as a data structure represented by a vector for each class to which the item consumed by the user belongs, Store in the user interest storage unit 2.

  The similarity calculation unit 30 acquires the user's interest (vector) from the user interest storage unit 2, measures the similarity of consumption patterns between users that are similar between two adjacent times (t, t-1), The consumption pattern is stored in the consumption pattern group storage unit 3.

  The transition probability calculation unit 40 reads a consumption pattern from the consumption pattern group storage unit 3, calculates a transition probability for each individual content or class in a similar consumption pattern group, and stores it in the transition probability storage unit 4.

  The predicted value calculation unit 50 reads the transition probability from the transition probability storage unit 4 and calculates the recommended value of the content for the individual.

  Hereinafter, the processing in the above configuration will be described in detail.

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

  Step 1) The time information adding unit 10 acquires the consumption history in the consumption history DB1 and the taxonomy from the information source, converts the time information of the consumption history into time, and gives the time to the consumption history. 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 consumption history, assuming that the first time on the consumption history is S and the value representing the period is D, each time t (i) is S + D × The time from i to S + D × (i + 1) is handled.

Step 2) The user interest construction unit 20 obtains time information and constructs the user's interest. A consumed item at a time t of the user v 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 user u manages it as a set of contents consumed at time t. A vector v _{u, i} is formed for each item i consumed by the user u. The vector indicates time, and the element in the column stores the consumption frequency of the item i at each time.

A vector v _{u, c} is formed for each class c to which the item consumed by the user u belongs. The column of the vector 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, only the set is checked, and if propagation is determined, the amount of calculation is reduced.

The value of the element corresponding to this vector column t is calculated as follows. f (c) is a function that returns a set of child classes of c, and c ^s is a child class of class c.

上記のようにして求められたベクトルをユーザ興味情報としてユーザ興味記憶部２に格納する。

The vector obtained as described above is stored in the user interest storage unit 2 as user interest information.

Step 3) The similarity calculation unit 30 reads a vector from the user interest storage unit 2 and extracts a consumption pattern of similar users. For the recommended user a, the consumption history up to the latest time x times from the current time T is particularly considered. Specifically, by particularly considering the history from T−x + 1 to T, it is possible to realize a recommendation that emphasizes the consumption history at the latest x time. For this reason, the vector va _{, c} is reduced to the dimension of the latest time. A consumption vector ra _{, c} relating to the class c in the consumption history at the latest x time is given. The consumption vector ra _{, c} is calculated as follows. In the present invention, since only t and t−1 are considered, x is 2.

  In the following formula (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.

次に、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を最終時刻とする類似度計算式を以下に示す。cos()はcosine類似度を示す。Ｃはユーザｕに対し、ステップ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。これは計算量削減のためである。

Next, a similarity calculation formula in which a certain time t between a user a and a user u for a certain class c is the final time is shown below. cos () indicates the cosine similarity. C indicates to the user u the class set calculated in step 2 and not all classes on the taxonomy. This is to reduce the calculation amount.

ユーザａとユーザｕの消費アイテムの類似度の計算は、消費履歴がスパースになりすぎることを避けるためと、計算量の削減のため、時系列を考慮しないものとする。消費アイテムの類似度(S_I(a,u))は、ユーザｕのクラスｃ配下のｔ−１とｔでのアイテム集合をＩ_ｕとすると、非特許文献2と同様に、以下の式(7)から計算する。式（7)において、jac()はjaccard係数を表す。なお、アイテムの類似度の計算方式は、以下の式(7)以外の公知の方法を用いてもよい。

The calculation of the similarity between the consumption items of the user a and the user u does not consider the time series in order to avoid the consumption history becoming too sparse and to reduce the calculation amount. The similarity (S _I (a, u)) of the consumption item is expressed by the following equation (1), where I _u is an item set at t−1 and t under the class c of the user u. Calculate from 7). In equation (7), jac () represents a jaccard coefficient. The item similarity calculation method may be a known method other than the following equation (7).

次に、ユーザａとユーザｕの消費パターン間の類似度(S_a,u)は、以下の式（8）により計算する。Ｎは正規化関数である。

Next, the similarity (S _{a, u} ) between the consumption patterns of the user a and the user u is calculated by the following equation (8). N is a normalization function.

上記式(8)により求められたユーザａ，ｕ間のパターン間の類似度（消費パターン類似度）(S_a,u)は消費パターン群記憶部3に格納する。

The similarity between patterns a between the users a and u (consumption pattern similarity) (S _{a, u} ) obtained by the above equation (8) is stored in the consumption pattern group storage unit 3.

Step 4) Next, the transition probability calculation unit 40 reads the consumption pattern similarity from the consumption pattern group storage unit 3 and normalizes the consumption pattern similarity between 0 and 1. Hereinafter, the normalized one is written as sa _{, u} .

  Then, the conditional probability of consuming item i (t-1) at time t-1 and consuming another item i (t) at time t is expressed by the following equation (9). Here, I (t−1) is an item set consumed at time t−1.

同じく、時刻ｔ−１にあるクラスc(t−1)を消費し、時刻ｔに別のクラスｃ(t）を消費した条件付き確率は式（10）のようになる。ここで、ｃ(t−1)は、時刻ｔ−１で消費されたクラス集合である。

Similarly, the conditional probability of consuming the class c (t−1) at the time t−1 and consuming another class c (t) at the time t is expressed by Equation (10). Here, c (t−1) is a class set consumed at time t−1.

一方、時刻ｔにおけるクラスｃ内のアイテムｉの生起確率は、時刻ｔ−１における生起頻度を受け継ぐとし、式(11)のように計算する。

On the other hand, the occurrence probability of the item i in the class c at the time t is assumed to inherit the occurrence frequency at the time t−1, and is calculated as in Expression (11).

同じく、時刻ｔにおける親クラスｐｃ内のクラスｃの生起確率ｔは、時刻ｔ−１におけるクラスｃの生起頻度を受け継ぐとし、式(12)のように計算する。

Similarly, the occurrence probability t of the class c in the parent class pc at the time t is calculated as the equation (12) assuming that the occurrence frequency of the class c at the time t−1 is inherited.

なお、タクソノミのルートクラス直下のクラスの生起確率は、一様に生起すると仮定する。直下のクラス数をＮとすると、生起確率は１／Ｎとなる。

It is assumed that the occurrence probability of the class immediately below the taxonomy root class occurs uniformly. When the number of classes immediately below is N, the occurrence probability is 1 / N.

  The occurrence probabilities obtained by the above formulas (9), (10), (11), and (12) are stored in the occurrence probability storage unit 4.

  Step 5) The predicted value calculation unit 50 reads the occurrence probability from the occurrence probability storage unit 4, and calculates the predicted value P of the item i at time t + 1.

  The predicted value P of item i at time (t + 1) is calculated as shown in equation (13).

If the item set consumed by the user a at time t is I _a and the class set is C _a ,

として計算する。

Calculate as

  In the above description, the transition probability is calculated using only the consumption history of t−1 and t. However, the transition probability can be updated using the consumption history of t + 1 in the future. In that case, update is performed using the following equation (14). In the following equation (14), α is a parameter that determines how much past history is important.

以下に、従来技術における非特許文献１の技術との違いを以下に示す。

Below, the difference with the technique of the nonpatent literature 1 in a prior art is shown below.

  (1) The present invention does not use all the consumption histories possessed by the user but uses only a part of the organizations in time series. For the recommended user, only the latest consumption history is used, and for the similarity calculation target user, only an arbitrary partial organization in time series is used. As a result, it is possible to cope with fluctuations in the interests of users, avoid sparse problems, and 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 be avoided.

  (2) As an item set for predicting an item, only the portion where the consumption pattern of the user is similar is used instead of using the similarity for each user. 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, differences from the technique of Non-Patent Document 2 will be shown.

  (1) Within a cube based on tensor decomposition, a user set having similar consumption behaviors at times t-1 and t is not managed, but is managed based on user similarity via a taxonomy class. The point of focus is the same as in Non-Patent Document 2 in that the recommendation can be personalized and the sparsity problem can be avoided. However, if the number of similar users is narrowed down, there is an advantage that the recommendation matching rate can be improved.

  (2) There is an advantage that recommended content can be presented based on systematized information of classes.

  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.

  The operation of each component of the content recommendation device shown in FIG. 1 can be constructed as a program and installed in a computer used as the content recommendation device, or can be distributed via a network.

  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 Consumption pattern group storage unit 4 Transition probability storage unit 10 Time information addition unit 20 User interest construction unit 30 Similarity calculation unit 40 Transition probability calculation unit 50 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;
For the user interest structured in the data of the user interest storage means, measure the consumption pattern of similar users between two adjacent times, and store the similarity in the consumption pattern group storage means,
In the consumption pattern group similar to the consumption pattern group storage means, the transition probability calculation means for calculating the transition probability for each individual of the content or class and storing it in the transition probability storage means;
A predicted value calculation means for obtaining each transition probability of the transition probability storage means and calculating a recommended value of content to an individual;
A content recommendation device comprising: The user interest building means is
Refer to the consumption history to which the time is given, and represent the number of items consumed by the user as a vector by dividing the number of consumptions at regular time intervals. Means for generating and storing in said user interest storage means;
The similarity measuring means includes
From the user-specific vector for each class of the user interest storage means, obtain the similarity between users of the number of times the item is consumed between the two most recent times of the current time, and the consumption pattern group storage means as consumption patterns of similar users Including means for storing in
The transition probability calculation means includes:
Means for obtaining a transition probability of consumption between times based on consumption patterns of items at time (t−1) and time t from the consumption pattern group storage means;
The predicted value calculation means includes
The content recommendation device according to claim 1, further comprising means for determining a recommended item by applying the transition probability to time (t + 1). 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 consumption history of the user's content having the evaluation value for the user's content and the evaluated time information, and the taxonomy of the content obtained from the information source. A time giving step for converting time information of the history into time and giving it to the consumption history;
A 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;
The similarity measurer measures the user's consumption pattern that is similar between two adjacent times for the user's interest structured in the data of the user interest storage device, and stores it in the consumption pattern group storage device Steps,
Transition probability calculating means calculates a transition probability for each individual of the content or class in the similar consumption pattern group of the consumption pattern group storage means and stores it in the transition probability storage means; and
A predicted value calculating means for acquiring each transition probability of the transition probability storage means and calculating a recommended value of content to an individual;
The content recommendation method characterized by performing. In the user interest building step,
Refer to the consumption history to which the time is given, and represent the number of items consumed by the user as a vector by dividing the number of consumptions at regular time intervals. Generated and stored in the user interest storage means,
In the similarity measurement step,
From the user-specific vector for each class of the user interest storage means, obtain the similarity between users of the number of times the item is consumed between the two most recent times of the current time, and the consumption pattern group storage means as consumption patterns of similar users Stored in
In the transition probability calculation step,
From the consumption pattern group storage means, based on the consumption pattern of the item at time (t-1) and time t, obtain the transition probability of consumption between times,
In the predicted value calculation step,
The content recommendation method according to claim 3, wherein a recommended item is determined by applying the transition probability at time (t + 1). Computer
The content recommendation program for functioning as each means of the content recommendation apparatus of Claim 1 or 2.

Images