JP2012022570A - Object recommendation apparatus, object recommendation method, object recommendation program and object recommendation system - Google Patents

Object recommendation apparatus, object recommendation method, object recommendation program and object recommendation system Download PDF

Info

Publication number
JP2012022570A
JP2012022570A JP2010160854A JP2010160854A JP2012022570A JP 2012022570 A JP2012022570 A JP 2012022570A JP 2010160854 A JP2010160854 A JP 2010160854A JP 2010160854 A JP2010160854 A JP 2010160854A JP 2012022570 A JP2012022570 A JP 2012022570A
Authority
JP
Japan
Prior art keywords
object
user
probability
target user
based
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
JP2010160854A
Other languages
Japanese (ja)
Other versions
JP5481295B2 (en
Inventor
Noriaki Kawamae
徳章 川前
Original Assignee
Ntt Comware Corp
エヌ・ティ・ティ・コムウェア株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ntt Comware Corp, エヌ・ティ・ティ・コムウェア株式会社 filed Critical Ntt Comware Corp
Priority to JP2010160854A priority Critical patent/JP5481295B2/en
Publication of JP2012022570A publication Critical patent/JP2012022570A/en
Application granted granted Critical
Publication of JP5481295B2 publication Critical patent/JP5481295B2/en
Application status is Active legal-status Critical
Anticipated expiration legal-status Critical

Links

Images

Abstract

It is possible to provide a novel object that is difficult for a user to notice.
A popularity calculation unit that recommends an object to a target user and calculates popularity information so that the larger the number of users who have performed a predetermined action on the object, the smaller the value is. 23, a degree-of-precedence calculation unit 21 that calculates the degree-of-precedence information indicating the degree to which another user has acted on the object in advance of the target user, and has elapsed since the target user has acted on the object Based on the importance level calculation unit 22 that calculates the importance level information indicating the importance level of the object for the target user based on the time, the other users can determine whether the other users are based on the popularity level information, the precedence level information, and the importance level information. An innovator probability calculation unit 24 that calculates an innovator probability based on the probability of acting on an object prior to the target user, and an object recommended to the target user based on the innovator probability It includes a recommendation object extracting unit 40 for extracting the.
[Selection] Figure 1

Description

  The present invention relates to an object recommendation device, an object recommendation method, an object recommendation program, and an object recommendation system.

  In recent years, online stores on the Internet have identified or identified individual Internet users (hereinafter referred to as users) and changed the services and contents provided to users according to their attributes and behaviors, so-called service Personalized. In the personalization of services, it is performed to find users who have similar preferences to the target user and recommend objects that those users prefer.

For example, in Patent Document 1, a user's access history is used to extract a relative relationship between the target user and another user, and the object included in the history data of the other user is a score obtained by quantifying the extracted relative relationship. A method of ranking is disclosed.
Further, in Non-Patent Document 1, in order to recommend an object that meets the latest preference of each user, not only the past purchase behavior is similar to the target user, but also the degree of purchase preceding the target user. A method for emphasizing a high history of other users is disclosed.

JP 2008-305055 A

Noriaki Kawamae, Akira Sakano, Takeshi Yamada, Nobuo Ueda "Collaborative Filtering Focusing on User's Preference and Time Precedence" IEICE Transactions D Vol. J92-D, NO. 6, pp. 767-776, 2009

  However, in Patent Document 1 and Non-Patent Document 1, there is a problem that an object that is generally not well known is not recommended because an object that is generally well known is easily recommended. In addition, since the well-known object is generally purchased as a standard object for a long time, there is a problem that it is not a novel object for the user.

  The present invention has been made in view of the above points, and an object of the present invention is to make it possible to provide a novel object that is difficult for the user to notice.

  The present invention has been made to solve the above-described problem, and is an object recommendation device that recommends an object to a target user, and as the number of other users who have performed a predetermined action on the object increases. A popularity degree calculation unit for calculating popularity degree information indicating the popularity degree of the object so as to take a small value, and a degree of the other user performing the action on the object in advance of the target user Precedence level calculation unit for calculating the priority level information, and importance level information indicating the level of importance of the object for the target user based on a time elapsed since the target user performed the action on the object. Based on the importance calculation unit to be calculated, the popularity information, the precedence information, and the importance information, the other user precedes the target user and the object. An innovator probability calculation unit that calculates an innovator probability based on the probability that the user has acted, and a recommended object extraction unit that extracts an object recommended to a target user based on the innovator probability. This is an object recommendation device.

  According to this invention, the object recommended to the target user is an object in which another user performs a predetermined action on the object prior to the target user, and the other user precedes the target user. The innovator probability based on the probability of performing a predetermined action on the user and the user transition probability that is the probability of performing the predetermined action on a predetermined object and then performing the action on another object are extracted. As a result, an object that has already performed a predetermined action by another user who has performed a predetermined action on the object prior to the target user, and an object that still has a small number of users is recommended. As a result, it is difficult for the user to notice himself / herself, and a novel object can be provided for each user.

  In addition, according to one aspect of the present invention, after each user performs the action on a predetermined object, the user takes the action on the predetermined object after performing the action on the predetermined object. A user transition probability calculating unit that calculates a probability of performing the action on another object as a user transition probability is further provided, and the recommended object extracting unit is based on the innovator probability and the user transition probability. It is characterized by extracting objects recommended to the user.

  According to this invention, the innovator probability is calculated based on the precedence information, the importance information, and the popularity information. As a result, an object that performs a predetermined action in common among users, and information that other users are performing a predetermined action in advance of the object, and target use of the object The recommended object can be extracted according to the importance level for the person and the popularity of the object. Therefore, the recommended object can be extracted depending on whether or not the object is important for the target user.

  Further, according to one aspect of the present invention, in the object recommendation device, the leading degree calculation unit includes a time when the target user performs a predetermined action on the object, and another user applies the action to the object. The leading degree information is calculated on the basis of the time when the object is released and the time when the object is disclosed.

  According to the present invention, the leading degree information is calculated based on the time when the target user performs a predetermined action on the object, the time when another user acts on the object, and the time when the object is released. ing. As a result, the preceding degree information of other users can be accurately calculated for the object, so that an appropriate object can be recommended to the target user.

  In addition, according to one aspect of the present invention, in the object recommendation device, the innovator probability calculation unit is configured so that each user performs each of the preceding degrees with the N users (N is a natural number) ahead of the target user. The innovator probability is calculated based on each probability based on the probability.

  According to the present invention, the innovator probability is calculated based on the respective probabilities based on the probabilities of the respective priorities in which the other user has performed N persons (N is a natural number) ahead of the target user. Yes. Accordingly, the innovator probability based on the probability that the other user has performed the action on the object in advance of the target user according to the number of persons preceding the target user and the preceding degree. Therefore, an appropriate object can be recommended to each target user.

Further, according to one aspect of the present invention, in the object recommendation device, the innovator probability calculation unit is based on a probability of each degree of advance that each user has performed the action one person ahead of the target user. When a matrix having each probability as an element is represented as P, the innovator probability is calculated based on PN .

According to this invention, based on PN , the innovator probability is calculated based on the probability that another user has acted on the object prior to the target user. Here, the (m, n) component of PN (where m is a natural number and n is a natural number) is a predecessor in which the user m performs a predetermined action on the object N times ahead of the target user n. It is a probability based on the probability of the degree. Accordingly, since it calculates the innovators probability based on the matrix P N, it is possible to recommend the appropriate object in the object user.

  Also, an aspect of the present invention is an object recommendation device that includes a popularity degree calculation unit, a leading degree calculation unit, an importance degree calculation unit, and a recommended object extraction unit, and recommends a digitized object to a target user. Is an object recommendation method executed by the popularity calculation unit that indicates the popularity of the object such that the popularity calculation unit takes a smaller value as the number of other users who have performed a predetermined action on the object increases. The degree of popularity calculation procedure for calculating degree information, and the degree of advancement in which the preceding degree calculation unit calculates the degree of advancement information indicating the degree to which the other user has acted on the object in advance of the target user. Based on the calculation procedure and the time that has passed since the target user performed the action on the object, the importance calculation unit determines the object for the target user. Based on the importance calculation procedure for calculating the importance information indicating the importance, the popularity information, the advance information, and the importance information, the other user precedes the target user, and An innovator probability calculation procedure for calculating an innovator probability based on the probability of performing the action on an object, and a recommended object extraction in which the recommended object extraction unit extracts an object recommended to a target user based on the innovator probability A method for recommending an object.

  Further, according to one aspect of the present invention, a computer as an object recommendation device that recommends an object to a target user takes a smaller value as the number of other users who have performed a predetermined action on the object increases. A first step of calculating popularity information indicating the popularity of the object, and first calculating priority information indicating a degree that the other user has acted on the object prior to the target user. And a third step of calculating importance information indicating importance of the object for the target user based on a time elapsed since the target user performed the action on the object. Based on the popularity information and the precedence information, the probability that the other user has performed the action on the object ahead of the target user. A fourth step of calculating the innovator probability brute, based on the innovators probability, an object recommended program for executing a fifth step of extracting the object to be recommended to the target user, the.

  Further, one aspect of the present invention is an object recommendation system that recommends an object to a target user, wherein the value increases as the number of other users who have performed a predetermined action on the object increases. A popularity degree calculation unit for calculating popularity degree information indicating the popularity degree of an object, and a leading degree for calculating the leading degree information indicating the degree that the other user has acted on the object in advance of the target user. A calculation unit; and an importance calculation unit that calculates importance information indicating importance of the object for the target user based on a time elapsed since the target user performed the action on the object; Based on the popularity information, the precedence information, and the importance information, the probability that the other user has performed the action on the object ahead of the target user. An innovator probability calculation unit that calculates an innovator probability based on the data, and a recommended object extraction device that extracts an object recommended to a target user based on the innovator probability. This is an object recommendation system.

  According to the present invention, it is possible to provide a novel object that is difficult for the user to notice.

It is a functional block diagram of the object recommendation apparatus in the 1st Embodiment of this invention. It is the figure which showed one example of the table memorize | stored in the log memory | storage part. It is a figure for demonstrating the object recommendation method which this invention proposes. It is the figure which showed distribution of the purchase user number for every release elapsed days of an object. It is a figure for demonstrating the probability of the leading degree of the user 1 with respect to the user 2 through a 3rd user. It is the figure which showed an example of the table in which the data of PIP and UFP memorize | stored in the calculation result memory | storage part were stored. It is the flowchart which showed the flow of the object recommendation process. It is the flowchart which showed the flow of the calculation process of PIP. It is the table which compared the accuracy of the recommendation top N specialized for an individual by each technique using the data set of a music, a video, Netflix, and a Query. It is the table which compared UC, IC, Gini coefficient, AE coefficient, and AD coefficient with each method using the data set of a music, a video, Netflix, and Query. It is a figure for demonstrating the calculation method of a Gini coefficient. It is a functional block diagram of the object recommendation system in the 2nd Embodiment of this invention.

<First Embodiment>
Hereinafter, a first embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a functional block diagram of an object recommendation device according to the first embodiment of the present invention.
The object recommendation device 1 includes a storage unit 10, a calculation unit 20, an input unit 31, a display unit 33, and a recommended object extraction unit 40.

The storage unit 10 includes a log storage unit 11 and a calculation result storage unit 12.
The log storage unit 11 stores a user id, which is a code for identifying a user, an object id, which is a code for identifying an object, and the time when each object is purchased in association with each other. Here, the object is a generic term for products, images, articles, and the like.

  FIG. 2 is a diagram illustrating an example of the table 50 stored in the log storage unit 11. In the table 40, the user id, the object id, and the time when the object is purchased are associated one-to-one.

  Returning to FIG. 1, the calculation result storage unit 12 stores the PIP (Personal Innovator Probability) calculated by the innovator probability calculation unit 24 and the UFP (Personal Innovator Probability) calculated by the user transition probability calculation unit 25. User Flow Probability (user transition probability) is stored.

  FIG. 3 is a diagram for explaining the object recommendation method according to the present embodiment. In the figure, there is a group of target users (users to be recommended) and people who have purchased the same object as the target user and who purchased the object earlier than the target user (innovator). It is shown.

  The right side of the figure shows the purchase transition of the object set purchased by the group of innovators. After purchasing the object 51, the group of innovators purchases the object 52 or the object 53, and then purchases the object 54. Object 54 is an object recently purchased by a group of innovators.

  In the present embodiment, the object recommendation device 1 searches one or more innovators corresponding to each user from the purchase history of each user's object, and notices the object that the innovator has purchased earlier. Extract as an object. The object recommendation device 1 extracts the object most recently purchased by the innovator from the extracted objects. This object is a novel object that is difficult for the target user to notice. The object recommendation device 1 recommends the object to the target user.

Returning to FIG. 1, the calculation unit 20 includes a precedence calculation unit 21, an importance calculation unit 22, a popularity calculation unit 23, an innovator probability calculation unit 24, and a user transition probability calculation unit 25. .
For each object of the object set C ab that is common between the users a and b, the leading degree calculation unit 21 determines that the user b (u b ) is the user a (u a ) for the object i (i is an integer of 0 or more). The leading degree r (i; b, a) is calculated.

  FIG. 4 is a diagram illustrating the distribution of the number of purchased users for each release elapsed day of the object. The horizontal axis in FIG. 4 represents the number of days elapsed from the release of the object until the purchase, and the vertical axis represents the logarithm of the number of users who purchased the object for each elapsed day. From the figure, since the logarithm of the distribution of the number of purchased users can be approximated to a straight line by a semilogarithmic graph, there is a peak in the number of purchased users immediately after the release of the object, and the number decreases exponentially. Therefore, the time distribution of the number of purchased users such as music and videos described later in the present embodiment can be approximated by an exponential distribution.

Therefore, the leading degree calculation unit 21 uses the exponential distribution to determine the degree r (i; b,) that the user b (u b ) precedes the user a (u a ) for the object i (i is an integer of 1 or more). a) is calculated by the following equation (1).

Here, t b, i is the time when user b purchased object i, t a, i is the time when user a purchased object i, T i is the release time of object i, and t i (overline) is the object. i is the average of all users at the time of purchase. The “(overline)” represents the average of the immediately preceding characters. If the user b has not purchased the object i prior to the user a, the preceding degree r is zero. In addition, the preceding degree r (i; b, a) increases as the user b purchases ahead of the user a.
The leading degree calculating unit 21 supplies the calculated leading degree r (i; b, a) to the innovator probability calculating unit 24.

Returning to FIG. 1, the importance calculation unit 22 calculates the importance w (i, a) of the object i with respect to the user a for each object in the object set Cab that is common between the users a and b. In general, since it is considered that the importance of an object for a user at the time of object purchase decreases exponentially, the importance calculation unit 22 calculates the importance w (i, a) by the following equation (2). To do.

Here, e a, i is the elapsed time from the time when the user a purchased the object i, and e a (overline) is the average of the elapsed time e a, i for all the objects i purchased by the user a. It is a thing. The importance w (i, a) of the object i with respect to the user a becomes higher as the elapsed time from the purchase of the object i is shorter. In other words, the importance w (i, a) of the object i with respect to the user a becomes higher as the purchase is made more recently.
The importance calculation unit 22 supplies the calculated importance w (i, a) to the innovator probability calculation unit 24.

The popularity calculation unit 23 calculates the popularity v (i) of the object i by the following equation (3) for each object of the object set Cab that is common between the users a and b.

Here, U i is the total number of users who purchased the object i. The popularity v (i) of the object i is so high that no other user has purchased it.
The popularity calculation unit 23 supplies the calculated popularity v (i) to the innovator probability calculation unit 24.

The innovator probability calculation unit 24 uses the preceding degree r (i; b, a), the importance w (i, a) of the object i with respect to the user a, and the popularity v (i) of the object i. Then, the preceding degree PID (u a , u b ) of the user b with respect to the user a is calculated by the following equation (4).

Here, the expression (4) is expressed as r (i; b, a) × w (i, a) × v (i) for all objects i belonging to the object set C ab common to the users a and b. It means calculating the sum.
Subsequently, the innovator probability calculation unit 24 calculates the probability p (b | a) of the preceding degree of the user b with respect to the user a using the PID, using the following equation (5).

Here, the denominator on the right side is obtained by adding the preceding degrees PID (u, u a ) of other users viewed from the user a to all users.
Subsequently, the innovator probability calculation unit 24 calculates the probability p (dot) (b, a) of the preceding degree of the user b with respect to the revised user a to implement the ergodic Markov chain by the following equation (6). calculate.

  Here, if no other user u has purchased an object prior to the user a (otherwise in equation (6)), the innovator probability calculation unit 24 determines that the user b with respect to the revised user a The probability of leading degree p (dot) (b, a) is assumed to be the reciprocal of the total number of users U.

  The innovator probability calculation unit 24 calculates the probability p (dot) (n | m) of the preceding degree of the user n with respect to the user m with a combination of all users. Then, the innovator probability calculation unit 24 calculates a matrix P (dot) in which the component of n rows and m columns is p (dot) (n | m).

  Subsequently, the innovator probability calculation unit 24 uses the matrix P (dot) to implement an ergodic Markov chain, and uses the matrix P (dot) for each of the preceding degrees purchased by each user ahead of the target user. The innovator probability matrix P (double dot) having each probability based on the probability p (dot) as an element is calculated by the following equation (7).

  Here, the innovator probability matrix P (double dot) is a probability p (double dot) (b, b) based on the probability p (b | a) of the advance degree that the user b purchases one user ahead of the user a. a) is an innovator probability matrix with (b, a) component of the matrix, α is a weight parameter (a number between 0 and 1), U is the total number of users, e is the length U of all elements 1 Column vector. According to this definition, the innovator probability matrix P (double dot) is configured as a probability matrix representing the probability of personal innovator transition. This innovator probability matrix P (double dot) converges to a steady distribution.

  Finally, the innovator probability calculation unit 24 calculates a PIP that takes into account the degree of precedence among users through a plurality of steps. For example, when three users (user 1, user 2, user 3) are assumed and a user focuses on a certain object, the user 3 precedes the user 2 with the object. Even if purchased, the user 1 purchased the object ahead of the user 3.

When a user purchases in advance through a small number of steps, there is a more direct relationship between the users, and the user who purchased in that case is considered to be a more important Innovator, The user's innovator probability needs to be raised relatively.
On the other hand, when a user purchases in advance through a large number of steps, there is less direct relationship between users, and the user who purchased in that case is not a very important Innovator. There is a need to relatively reduce the user's innovator probability.
Therefore, the innovator probability calculation unit 24 uses, as an example, a matrix P (overline) whose element is an innovator probability based on the probability that another user has purchased the object in advance of the target user. ).

Here, β is a parameter for adjusting the effect of the leading degree, and is a value from 0 to 1, for example. P (double dot) N (N is a natural number) is a matrix whose components are each probability based on each probability that each user purchases N users (N is a natural number) ahead of the target user. Also, the larger the number of N, the more 1 / N! Therefore, the effect of P (double dot) N is reduced. Further, exp (−β) is a normalization coefficient.

  The second item in parentheses on the right side is a term representing the degree of precedence of the user m with respect to the user n and represents the effect of the degree of precedence with one user in between. Similarly, the N-1 item in the parenthesis on the right side is a term representing the degree of precedence of the user m with respect to the user n, and represents the effect of the degree of precedence with (N-2) users in between.

  Here, what is meant by the second item in parentheses on the right side of Equation (8) will be described. FIG. 5 is a diagram for explaining the probability of the preceding degree of the user 1 with respect to the user 2 through the third user. In the figure, it is assumed that there are three users (user 1, user 2, user 3).

  In this case, the purchase pattern includes a first pattern in which objects are purchased in the order of user 1, user 1 and user 2, a second pattern in which objects are purchased in order of user 1, user 2 and user 2, and user 1, There is a third pattern in which objects are purchased in the order of user 3 and user 2.

In the first pattern in the figure, the probability of purchasing an object prior to the user 1 of the user 1 (hereinafter referred to as the advance purchase probability) is p (double dot) (1, 1), and the user 2 of the user 1 The prior purchase probability for user 2 is p (double dot) (1, 2), the previous purchase probability for user 2 is p (double dot) (2, 2), and the previous purchase probability for user 3 for user 1 is p. In (double dot) (1, 3), the prior purchase probability of the user 3 for the user 2 is p (double dot) (3, 2). Then, in FIG. 5, the user 1 precedes the user 2 and the prior purchase probability (P (double dot)) 2 (1,2) precedes the user 2 by the following formula (9).

This advance purchase probability (P (double dot)) 2 (1,2) is the same as the component in the first row and second column of (P (double dot)) 2 represented by the following equation (10).

Therefore, the second item in parentheses on the right side of the mathematical formula (8) is a term representing the degree of precedence of the user m with respect to the user n and represents the effect of the degree of precedence with one user in between.
Further, the meaning of the (N−1) -th item in parentheses on the right side of Expression (8) will be described. When a matrix having a probability p (double dot) (m, n) based on the probability of the preceding degree purchased by the user m one prior to the target user n is represented as P (double dot), P (Double dot) N-1 is a matrix having each probability as an element based on the probability of each preceding degree purchased by each user preceding the target user by N-1 (N is a natural number).
Returning to FIG. 1, the innovator probability calculation unit 24 generates a matrix P (overline) having an innovator probability based on the probability that another calculated user has purchased the object ahead of the target user. The calculation result storage unit 12 stores the result as PIP.

Subsequently, the user transition probability calculation unit 25 calculates UFP by the following method. The purpose of UFP is to give each object an appropriate weight in order to estimate the number of consumers who have purchased another object after the purchase of one particular object. In order to calculate the transition probability p ab , after the purchase of the object a, a function relating to the consumer who purchased the object b at time t is considered. Depending on this probability, it can be predicted that the object will be purchased next.

In the first embodiment of the present invention, a transition probability between objects is constructed using a continuous-time Markov process that satisfies a Markov characteristic and takes a value obtained from one set of state space. A continuous-time Markov process is most simply defined by specifying a transition probability q ij . This process is typically given as the ij component of all transition probability matrices Q for Markov chain transitions.

  In order to preserve this probability, the non-diagonal element of the transition probability matrix Q is not negative, and the diagonal element called the jump rate must satisfy the following equation (11).

Here, q i is a chain probability from the object i. In the statistical process, the probability p ij of transition from the object i to the object j is independent of the time of staying in the object i, and must satisfy p ii = 0 and Σp ij = 1 (where j ≠ i). .
Accordingly, the probability p ij of transition from the object i to the object j is expressed by the following formula (12).

The period of staying at the object i follows an exponential function with the transition probability q i as a coefficient. From the characteristic of the random variable of the exponential distribution, E (t i ) that is the average of the times t u and ij that the user u took from the object i to the purchase of the object j is expressed by the following equation (13) using the transition probability q i : ).

The user transition probability calculation unit 25 calculates the transition probability q i according to the equation (13), and then calculates the probability p ij of transition from the object i to the object j according to the following equation (14).

From equation (12), the transition probability q ij is expressed by the following equation (15).

Subsequently, in order to predict a trend at time t = t f , the user's transition probability u (j | i, t from object i to another object in period [0, t f ], which is the ij component of matrix U, f ) is expressed by the following equation (16).

Here, P (t) represents a transition matrix that is the ij component p ij . Formally, if the state space is finite, the transition probability is estimated by the following differential equation (17) from equation (12).

  Here, I is an identity matrix. The solution of the differential equation (17) is expressed by the following equation (18).

Subsequently, assuming that Q is diagonalized by Q = MDM −1, the transition matrix function P (t) satisfying the Kolmogorov forward and back equation (Kolmogorov forward and backing equation) is expressed by the following equation (19). expressed.

  Since Q is large, the transition matrix function P (t) is approximated by the following equation (20) using Taylor approximation.

Finally, UFP is composed of u (i b | i a , ∞) and is expressed by the following equation (21).

  Accordingly, the user transition probability calculation unit 25 calculates UFP using the above equation (21). Then, the user transition probability calculation unit 25 stores the calculated UFP in the calculation result storage unit 12.

  FIG. 6 is a diagram illustrating an example of a table in which PIP and UFP data stored in the calculation result storage unit 12 are stored. FIG. 6A is a diagram illustrating an example of a table 61 in which PIP data stored in the calculation result storage unit 12 is stored. In the figure, in the table 61, the user id of the target user, the user id of another user, and the personal revolutionary probability PIP of the other user with respect to the target user are associated with each other.

  FIG. 6B is a diagram illustrating an example of a table 62 in which UFP data stored in the calculation result storage unit 12 is stored. In the figure, in the table 62, the object id of the original object, the next object id, and the probability UFP of transition from the original object to the next object are associated.

Returning to FIG. 1, based on the input input from the input unit 31 from the outside, the user id of the target user is supplied to the recommended object extraction unit 32.
The recommended object extraction unit 32 reads the PIP and UFP corresponding to the user id of the target user supplied from the input unit 31 from the calculation result storage unit 12. Then, the recommended object extraction unit 32 performs CPI (Continuous Probability of the Item Given Each User) corresponding to the user id of the target user based on the read PIP and UFP, and each user purchases an object. Probability).

The probability p (i b | u a ) that the user a (u a ) purchases the object b (i b ) using the PIP calculated by the equation (10) and the UFP calculated by the equation (21) is It is represented by the following formula (22).

Here, δ (i a | u j ) is 1 if the user j has purchased the object a (i a ) at the time of calculation of Expression (22), and 0 otherwise.
If the PIP of the personal innovator who purchased the object b (i b ) is high and the UFP to the object i b is high, the probability p (i b | that the user a (u a ) purchases the object b (i b ) u a ) becomes higher.

The recommended object extraction unit 40 calculates the probability p (i | u a ) for all the objects i. Then, the recommended object extraction unit 40 extracts N (referred to as top N) objects in descending order of the probability p (i | u a ).

  When an object purchased by the target user is included in the object list including the top N objects, the recommended object extraction unit 40 deletes the object from the object list and removes the deleted object. Is displayed on the display unit 33.

The recommended object extraction unit 40 extracts the top N objects and displays the object list including the extracted objects on the display unit 33. However, the present invention is not limited to this, and the probability p (i | u a ) is the highest. A high object may be extracted, and an object corresponding to the extracted object may be displayed on the display unit 33.

  FIG. 8 is a flowchart showing the flow of the PIP calculation process. First, the calculation unit 20 reads a log from the storage unit 10 (step S101). Next, the innovator probability calculation unit 24 calculates a PIP based on the read log, and stores the calculated PIP in the calculation result storage unit 12 (step S102). Next, the user transition probability calculation unit 25 calculates a UFP based on the read log, and stores the calculated UFP in the calculation result storage unit 12 (step S103).

  Next, the recommended object extraction unit 40 reads the PIP and UFP corresponding to the user id of the target user supplied from the input unit from the calculation result storage unit 12. Then, the recommended object extraction unit 40 calculates CPI based on the read PIP and UFP (step S104). Next, the recommended object extraction unit 40 causes the display unit 33 to display an object list in which objects are arranged in descending order of CPI (step S105). Above, the process of this flowchart is complete | finished.

  FIG. 8 is a flowchart showing the flow of the PIP calculation process. First, the leading degree calculation unit 21 calculates leading degree information r from the log read from the log storage unit 11 and supplies it to the innovator probability calculation unit 24 (step S201). The importance calculation unit 22 calculates importance information w from the log read from the log storage unit 11, and supplies the importance information w to the innovator probability calculation unit 24 (step S202). The popularity calculation unit 23 calculates popularity information i from the log read from the log storage unit 11, and supplies the popularity information i to the innovator probability calculation unit 24 (step S203).

  The innovator probability calculating unit 24 calculates a leading probability PID based on the supplied leading degree information r, importance degree information w, and popularity degree information i (step S204). The innovator probability calculation unit 24 calculates PIP based on the calculated preceding probability PID and stores it in the calculation result storage unit 12 (step S205). Above, the process of this flowchart is complete | finished.

<Effect of the present invention>
Next, the results of experiments conducted to verify the effects of the method proposed by the present invention will be described.
<About the data set used in the experiment>
In the first embodiment of the present invention, experimental results for four data sets will be described. Two of the four data sets are purchase histories obtained from a music or video download service at an online store in Japan. The third is a movie review history Netflix used as evaluation data for a collaborative filtering contest. The last one is a search query log. Hereinafter, each data set will be described.

First, the details of the online store purchase history are as follows. The purchase history of the music download is from April 1, 2005 to July 31, 2006, and includes 44,527 objects purchased by 84,620 users.
Each purchase history includes a user ID, the title of the purchased song, the artist name, the title of the CD album, the purchase date and time, and the price.

The purchase history of video downloads is from September 1, 2005 to February 28, 2006, and includes 4,064 objects purchased by 7,537 users.
Each purchase history includes a user ID, a purchased video title, a director name, a purchase date and time, and a price.

  Most objects of these data are newly released, and are first made available for purchase on the online site as CDs or videos (DVDs) immediately after release at the store.

  Netflix is 100,480,507 evaluation data for 17,770 movies evaluated by 480,189 users from November 11, 1999 to December 31, 2005. Users who have rated at least 20 movies are selected. Also, movies rated by at least 100 users are selected. This reduces the data size of the data set, resulting in 85,730,203 evaluation data for 9,264 movies evaluated by 136,589 users.

Each evaluation data includes a movie title ID, a user ID, an evaluation, and an evaluation date. Unlike the above two online store purchase histories, Netflix is a user's movie evaluation log, and its data is represented by a 5-level rating from 1 to 5.
In the collaborative filtering proposed in the present invention, purchase is predicted instead of user evaluation. Therefore, Netflix data was used as follows.

  The user's rating was converted from a five-level rating to a binary value from 0 to 1. That is, when the user has evaluated an object, the evaluation is 1 (meaning purchase), and otherwise 0 (means not purchased). Netflix (o) and Netflix (p) are calculated in the same way as the literature (N. Kawamae, H. Sakano, T. Yamada, Personalized recommendation based on the personal innovator degree. In ACM Recsys, pages 329-332, 2009). Was used.

  Next, the search query is made from a search engine log from April 1, 2006 to May 31, 2006. This data set is made up of 35, 325 and 842 negative elie histories. Each search query includes a query keyword ID, a user ID, and a history date / time.

<Experimental plan>
The purpose of the experiment is to predict which music, video, movie (query) the user will purchase (search) in the future from the user's past purchase history. In order to evaluate the prediction accuracy of recommendation, a simulation was performed in which the data set was randomly divided into K sub-data sets and cross-checked K times.

In order to test the model among the K sub-data sets, one sub-data set is used as evaluation data, and the other K-1 sub-data sets are used as training data.
This process was repeated K times, and each sub-data set was evaluated data only once. Each sub-data set is divided into two periods (learning period and test period), the sub-data set during the test period is referred to as test data, and the sub-data set during the learning period is referred to as learning data. Here, K is set to 10 as an example.

  The object recommendation device 1 ranks the remaining K-1 data objects using the proposed method, with all users included in the test period of one data set out of K divided as target users. The object recommendation device 1 presented the top N objects to the target user based on the ranking.

  Then, the object recommendation device 1 calculates the ratio of the presented objects included in the purchase history during the test period of the target user as the top N accuracy. Here, the accuracy of the top N is an index generally used to evaluate the prediction performance of collaborative filtering.

Here, the data set was applied to nine conventional recommendation methods, and the accuracy of the recommendation of the top N (N is 1, 5, or 10) was compared.
FIG. 9 is a table in which the accuracy of the top N of personalized recommendations is compared by each method using data sets of music, video, Netflix, and Query. When the t test differs significantly from p <0.05 and p <0.01 compared to all other methods, “*” and “**” are respectively marked.

Since one of the conventional recommendation methods, Popular recommends the most popular objects in the last month of the learning period, the recommended objects are not individualized for each user.
One of the conventional recommendation methods, Pearson and Cosine, is based on the user's similarity measured by Pearson's correlation coefficient or cosine similarity, respectively.

On the other hand, Item was measured by Pearson's correlation coefficient proposed in the literature (JKB Sarwa, G. Karypis and J. Riedl. Item-based collaborative filtering recommendation algorithms. In WWW, pages 285-295, 2001.) Based on content similarity.
bPLSA is based on PLSA (Probabilistic Latin Semantic Analysis) using Bernoulli distribution (T. Hofmann. Collaborative filtering via Gaussian probabilistic latent semantic analysis. In ACM SIGIR, pages 259-266, 2003).

  MEA is the Maximum Entropy Approach (maximum entropy) proposed in the literature (D. Pavlov and D. Pennock. A maximum entropy approach to collaborative filtering in dynamic, sparse, high-dimensional domains. In NIPS, pages 1441-1448, 2002). Law).

IFD is a collaborative filtering for latent feedback datasets proposed in the literature (Y. Hu, Y. Koren, and C. Volinsky. Collaborative filtering for implicit feedback datasets. In ICDM, pages 263-272, 2008.). is there.
EABIF is an early adaptation based information flow proposed in the literature (X. Song, C. Lin, B. Tseng, and M. Sun. Personalized recommendation driven by information flow. In ACM SIGIR, pages 509-516, 2006.). (EABIF) method.

  PID is PID calculated by Formula (4). PIP is a PIP calculated by Expression (8). PIP + UFP is a method proposed by the present invention. In this embodiment, α in equation (7) is set to 1, β in equation (8) is set to 1, and N is set to 5.

In FIG. 9, when the data set is music, video, Netflix (h) or Netflix (p), the PIP + UFP proposed by the present invention has the accuracy of the top 10 recommended objects as compared to all other methods. high.
That is, PIP + UFP indicates that the object to be purchased by the user in the future is estimated with higher accuracy than all other methods.

  This result can be explained by the feature that the objects included in the data set are composed of expensive objects. That is, it is considered that each user purchases these expensive objects motivated by preference rather than from necessity.

  PIP + UFP uses a time decay factor that increases the weight for an object purchased more recently, and distinguishes between users with the same preference. PIP + UFP also introduces object trends into object rankings. As a result, PIP + UFP, which is a method proposed by the present invention, ranks objects that match the user's latest preference at the top, so that the performance of object recommendation accuracy is improved.

  FIG. 10 is a table in which UC, IC, Gini coefficient, AE coefficient, and AD coefficient are compared by each method using data sets of music, video, Netflix, and Query. When the t test differs significantly from p <0.05 and p <0.01 compared to all other methods, “*” and “**” are respectively marked.

  In the figure, in addition to the top N accuracy mentioned above, the two coverage ratios (IC, UC), namely the Gini coefficient, the average time from release of the object (AE) or the average of expected purchase time (AD) Used to evaluate the performance of each of the above methods.

  Subsequently, each index will be described in detail. First, UC (User Coverage: coverage ratio of recommended users) is a ratio of the number of users who can recommend each recommendation method to the number of users who purchased the object during the test period. Since the object can be recommended to many users as the UC is higher, the system is more valuable for the entire user.

  On the other hand, IC (Item Coverage: coverage ratio of recommended objects) is the ratio of the number of titles that can be recommended by each recommendation method to the number of titles of objects purchased during the test period. The IC is an index indicating the size of the object domain in the system that can be recommended by the system. Therefore, a system with a low IC can present only a limited number of selected objects, and is therefore a low-value system for the user.

  The Gini coefficient is an index indicating the statistical dispersion of the distribution of the number of purchasers of the user with respect to the object. FIG. 11 is a diagram for explaining a Gini coefficient calculation method. The Gini coefficient g is a ratio of the area A surrounded by the 45 degree line 71 and the Lorenz curve 72 to the area (A + B) surrounded by the 45 degree line 71, the horizontal axis 73, and the vertical axis 74, In the formula, g = A / (A + B). The Gini coefficient takes a value from 0 to 1, and the closer the value is to 0, the smaller the difference in the number of purchased users for each object, and the closer the value is to 1, the larger the difference.

  If the Gini coefficient is 0, the distribution is completely equal, that is, all objects have been purchased by the exact same number of users. On the other hand, when the Gini coefficient is 1, the distribution is completely unequal, that is, one object is purchased by all users and the other objects are not purchased by any user.

  The result of a high Gini coefficient means that a few specific objects tend to be highly ranked by most users, only specific objects tend to be recommended, The difference between the recommended objects is reduced. That is, object recommendation means that it is not specialized for each user. On the other hand, the closer the Gini coefficient is to 0, the more specific the object recommendation is for each user, and the better the object recommendation is.

AE (Average Elapsed time) is an average of the elapsed time from the release of the object to the purchase. The smaller this value, the higher the novelty of the object.
AD (Average Difference time) is an average of the difference between the start time of the test period and the time of object purchase. If this value is large, the object becomes harder to notice.

In FIG. 10, when the data set is music, video, Netflix (p), the PIP + UFP proposed by the present invention has a higher IC, lower Gini coefficient, shorter AE, and AD than all other methods. Is long.
In other words, PIP + UFP can recommend many objects compared to all other methods, recommend object specific to each user, recommend highly novel objects, and recommend objects that are difficult to notice. It shows that.

In addition, when the data set is Netflix (h), the PIP + UFP proposed by the present invention has a higher IC, lower Gini coefficient, and shorter AE than the PIP + UFP proposed by the present invention. .
That is, PIP + UFP indicates that more objects can be recommended than all other methods, the object recommendation is specialized for each user, and the highly novel object is recommended.

Further, when the data set is Query, the PIP + UFP proposed by the present invention has a lower Gini coefficient, a shorter AE, and a longer AD than the PIP + UFP proposed by the present invention.
That is, PIP + UFP indicates that object recommendation is specialized for each user, highly novel objects are recommended, and objects that are difficult to notice are recommended compared to all other methods.

  From the results of FIG. 10, when PIP + UFP proposed in the present invention is used, the highest UC and IC are obtained, the Gini coefficient is the lowest (close to 0), and AD is the longest. On the other hand, from the result of FIG. 9, even when the PIP + UFP proposed in the present invention is used, the accuracy of the top N recommendations is not significantly improved. From this result, the PIP + UFP proposed in the present invention ranks and recommends different objects with little bias for each object compared to other methods.

  In fact, the conventional method that does not take into account the user's preference dynamics and the relationship between users has a low IC and a high Gini coefficient. These traditional methods are generally popular because they treat users who purchase the object later as similar users and use their logs as well as those who purchase the object immediately after release. , Rank trivial objects higher.

  With respect to the coverage ratio (UC) of recommended users, the conventional method considering the dynamics of the relationship between users takes a slightly lower value than other methods. The PIP + UFP proposed in the present invention can recommend an object that has not been purchased so far by most users as a novel object for the user.

  As described above, the object recommendation apparatus 1 according to the first embodiment of the present invention can recommend an object that is specialized for each user and has high novelty, and that is difficult to notice.

<Second Embodiment>
Next, a second embodiment of the present invention will be described. FIG. 12 is a functional block diagram of an object recommendation system according to the second embodiment of the present invention. Elements common to those in FIG. 1 are denoted by the same reference numerals, and detailed description thereof is omitted.
The object recommendation system 101 includes a storage device 110, a calculation device 120, a terminal device 130, and a recommended object extraction device 140.

  The storage device 110 includes a log storage unit 11 and a calculation result storage unit 12. The calculation device 120 includes a leading degree calculation unit 21, an importance degree calculation unit 22, a popularity degree calculation unit 23, an innovator probability calculation unit 24, and a user transition probability calculation unit 25.

  The terminal device 130 includes an input unit 31 and a display unit 33. The input unit 31 supplies information indicating the user id to the recommended object extraction device 140 based on an input signal supplied from the outside.

  The recommended object extraction device 140 reads the PIP and UFP corresponding to the user id supplied from the input unit 31 from the calculation result storage unit 12. Then, the recommended object extraction device 140 calculates the CPI by the above method based on the read PIP and UFP. Then, the recommended object extraction device 140 causes the display unit 33 to display N objects (N is an integer of 1 or more) in descending order of CPI. Since the processing flow is the same as in the first embodiment, a description thereof will be omitted.

  As described above, the object recommendation system 101 according to the second exemplary embodiment of the present invention can recommend an object that is specialized for each user and has high novelty and is difficult to notice.

  In the first embodiment or the second embodiment of the present invention, the case where an object is purchased as a user's action has been described. However, the present invention is not limited to this, and the object displayed on the display unit 33 (for example, The present invention can also be applied to the case of browsing articles. That is, the present invention can also be applied to an apparatus and system that recommends a novel object that is difficult for the user to notice based on the browsing history of the user browsing the object.

In view of the above points, the present invention is not limited to when a user purchases an object, and the user performs predetermined actions (for example, article browsing, image selection, etc.) on the object. It is also applicable when
In the present invention, the object is not limited to digitized music, electronic books, moving images, and the like, and the object includes books, furniture, and the like that are not digitized.

  In addition, you may make it implement | achieve a part of the object recommendation apparatus 1 which is the 1st Embodiment of this invention, for example, the calculation part 20, or the recommendation object extraction part 40 with a computer. In this case, the object detection program for realizing the function may be recorded on a computer-readable recording medium, and the object recommendation program recorded on the recording medium may be read into the computer system and executed. Good. Here, the “computer system” includes an OS (Operating System) and hardware of peripheral devices. The “computer-readable recording medium” refers to a portable recording medium such as a flexible disk, a magneto-optical disk, an optical disk, and a memory card, and a storage device such as a hard disk built in the computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in the computer system that becomes a server or a client in that case may be included and a program that holds a program for a certain period of time may be included. Further, the above program may be for realizing a part of the functions described above, or may be realized by a combination with the program already recorded in the computer system. .

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, the concrete structure is not restricted to this embodiment, The design etc. of the range which does not deviate from the summary of this invention are included.

DESCRIPTION OF SYMBOLS 1 Object recommendation apparatus 10 Memory | storage part 11 Log memory | storage part 12 Calculation result memory | storage part 20 Calculation part 21 Advancement degree calculation part 22 Importance degree calculation part 23 Popularity degree calculation part 24 Innovator probability calculation part 25 User transition probability calculation part 31 Input part 33 Display unit 40 Recommended object extraction unit 101 Object recommendation system 110 Storage device 120 Calculation device 130 Terminal device 140 Recommended object extraction device

Claims (8)

  1. An object recommendation device for recommending an object to a target user,
    A popularity calculation unit that calculates popularity information indicating the popularity of the object, so that the smaller the number of other users who have performed a predetermined action on the object, the smaller the value.
    A predecessor calculation unit for calculating a predecessor degree information indicating a degree of the other user performing the action on the object in advance of the target user; and elapses after the target user performs the action on the object. An importance calculating unit that calculates importance information indicating the importance of the object for the target user based on the calculated time;
    Innovation that calculates an innovator probability based on a probability that the other user has performed the action on the object in advance of the target user based on the popularity information, the precedence information, and the importance information. Person probability calculation unit,
    Based on the innovator probability, a recommended object extraction unit that extracts an object recommended to the target user;
    An object recommendation device comprising:
  2. Probability that each user performs the action on the other object after performing the action on the predetermined object based on the time required until the user takes the action on the other object after the user performs the action on the predetermined object. Is further provided with a user transition probability calculating unit that calculates the user transition probability,
    The object recommendation device according to claim 1, wherein the recommended object extraction unit extracts an object recommended to a target user based on the innovator probability and the user transition probability.
  3.   Based on the time when the target user performs a predetermined action on the object, the time when another user performs the action on the object, and the time when the object is released, The object recommendation apparatus according to claim 1, wherein the leading degree information is calculated.
  4.   The innovator probability calculation unit calculates the innovator probability based on each probability based on the probability of each preceding degree in which each user has performed the N users (N is a natural number) ahead of the target user. The object recommendation device according to any one of claims 1 to 3, wherein
  5. The innovator probability calculation unit represents P N as a matrix having each probability as a factor based on the probability of each preceding degree in which each user has performed the action one preceding the target user. 5. The object recommendation device according to claim 4, wherein the innovator probability is calculated based on the object.
  6. An object recommendation method executed by an object recommendation device that includes a popularity calculation unit, a precedence calculation unit, an importance calculation unit, and a recommended object extraction unit, and recommends an object to a target user,
    A popularity calculation procedure for calculating popularity information indicating the popularity of the object so that the popularity calculation unit takes a smaller value as the number of other users who have performed a predetermined action on the object increases. ,
    A leading degree calculation procedure in which the leading degree calculating unit calculates leading degree information indicating a degree of the other user performing the action on the object in advance of the target user;
    The importance calculation unit calculates importance information indicating importance of the object for the target user based on a time elapsed since the target user performed the action on the object. Procedure and
    Innovation that calculates an innovator probability based on a probability that the other user has performed the action on the object in advance of the target user based on the popularity information, the precedence information, and the importance information. Person probability calculation procedure,
    The recommended object extraction unit extracts a recommended object to be recommended to a target user based on the innovator probability; and
    An object recommendation method characterized by comprising:
  7. To a computer as an object recommendation device that recommends objects to target users,
    A first step of calculating popularity information indicating the popularity of the object so as to take a smaller value as the number of other users who have performed a predetermined action on the object increases,
    A second step of calculating preceding degree information indicating a degree of the other user performing the action on the object in advance of the target user;
    A third step of calculating importance information indicating importance of the object for the target user based on a time elapsed since the target user performed the action on the object;
    Based on the popularity information, the precedence information, and the importance information, an innovator probability is calculated based on a probability that the other user has acted on the object in advance of the target user. 4 steps,
    A fifth step of extracting objects recommended to the target user based on the innovator probability;
    Object recommendation program to execute
  8. An object recommendation system for recommending an object to a target user,
    A popularity calculation unit that calculates popularity information indicating the popularity of the object, so that the smaller the number of other users who have performed a predetermined action on the object, the smaller the value.
    A leading degree calculating unit that calculates leading degree information indicating a degree of the other user performing the action on the object in advance of the target user;
    An importance calculating unit that calculates importance information indicating importance of the object for the target user based on a time elapsed since the target user performed the action on the object;
    Innovation that calculates an innovator probability based on a probability that the other user has performed the action on the object in advance of the target user based on the popularity information, the precedence information, and the importance information. Person probability calculation unit,
    A computing device comprising:
    Based on the innovator probability, a recommended object extracting device that extracts an object recommended to the target user;
    An object recommendation system comprising:
JP2010160854A 2010-07-15 2010-07-15 Object recommendation device, object recommendation method, object recommendation program, and object recommendation system Active JP5481295B2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2010160854A JP5481295B2 (en) 2010-07-15 2010-07-15 Object recommendation device, object recommendation method, object recommendation program, and object recommendation system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP2010160854A JP5481295B2 (en) 2010-07-15 2010-07-15 Object recommendation device, object recommendation method, object recommendation program, and object recommendation system

Publications (2)

Publication Number Publication Date
JP2012022570A true JP2012022570A (en) 2012-02-02
JP5481295B2 JP5481295B2 (en) 2014-04-23

Family

ID=45776800

Family Applications (1)

Application Number Title Priority Date Filing Date
JP2010160854A Active JP5481295B2 (en) 2010-07-15 2010-07-15 Object recommendation device, object recommendation method, object recommendation program, and object recommendation system

Country Status (1)

Country Link
JP (1) JP5481295B2 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103700007A (en) * 2013-12-18 2014-04-02 东软集团股份有限公司 Recommendation method and system based on object
US10469826B2 (en) 2014-08-08 2019-11-05 Samsung Electronics Co., Ltd. Method and apparatus for environmental profile generation

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
CSNG200700789003; 川前 徳章、外2名: 'Innovatorの発見によるリコメンデーションのパーソナライズ手法の提案' 電子情報通信学会技術研究報告 第107巻,第78号, 20070524, p.19-24, 社団法人電子情報通信学会 *
CSNG200700976003; 市川 裕介: '利用者の好みをとらえ活かす-嗜好抽出技術の最前線-' 情報処理 第48巻,第9号, 20070915, p.972-978, 社団法人情報処理学会 *
CSNG200900279007; 川前 徳章、外3名: 'ユーザの嗜好の時系列性と先行性に着目した協調フィルタリング' 電子情報通信学会論文誌 第J92-D巻,第6号, 20090601, p.767-776, 社団法人電子情報通信学会 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103700007A (en) * 2013-12-18 2014-04-02 东软集团股份有限公司 Recommendation method and system based on object
US10469826B2 (en) 2014-08-08 2019-11-05 Samsung Electronics Co., Ltd. Method and apparatus for environmental profile generation

Also Published As

Publication number Publication date
JP5481295B2 (en) 2014-04-23

Similar Documents

Publication Publication Date Title
Jeon et al. A framework to predict the quality of answers with non-textual features
Lü et al. Recommender systems
Panniello et al. Experimental comparison of pre-vs. post-filtering approaches in context-aware recommender systems
Bozzon et al. Choosing the right crowd: expert finding in social networks
US7685091B2 (en) System and method for online information analysis
US8566256B2 (en) Universal system and method for representing and predicting human behavior
Li et al. Routing questions to appropriate answerers in community question answering services
US8095523B2 (en) Method and apparatus for context-based content recommendation
Liu et al. Online evolutionary collaborative filtering
Zhao et al. Improved recommendation based on collaborative tagging behaviors
Deveaud et al. Accurate and effective latent concept modeling for ad hoc information retrieval
Vig et al. Tagsplanations: explaining recommendations using tags
Yin et al. Challenging the long tail recommendation
Zhen et al. TagiCoFi: tag informed collaborative filtering
Kim et al. Collaborative filtering based on collaborative tagging for enhancing the quality of recommendation
US9275001B1 (en) Updating personal content streams based on feedback
US20100161385A1 (en) Method and System for Content Based Demographics Prediction for Websites
JP5662961B2 (en) Review processing method and system
US20090019034A1 (en) Media discovery and playlist generation
US8909626B2 (en) Determining user preference of items based on user ratings and user features
US9117236B1 (en) Establishing communication based on item interest
US8301623B2 (en) Probabilistic recommendation system
White et al. Predicting short-term interests using activity-based search context
Shi et al. Getjar mobile application recommendations with very sparse datasets
JP5736469B2 (en) Search keyword recommendation based on user intention

Legal Events

Date Code Title Description
A621 Written request for application examination

Free format text: JAPANESE INTERMEDIATE CODE: A621

Effective date: 20130220

A977 Report on retrieval

Free format text: JAPANESE INTERMEDIATE CODE: A971007

Effective date: 20131224

TRDD Decision of grant or rejection written
A01 Written decision to grant a patent or to grant a registration (utility model)

Free format text: JAPANESE INTERMEDIATE CODE: A01

Effective date: 20140121

A61 First payment of annual fees (during grant procedure)

Free format text: JAPANESE INTERMEDIATE CODE: A61

Effective date: 20140217

R150 Certificate of patent or registration of utility model

Ref document number: 5481295

Country of ref document: JP

Free format text: JAPANESE INTERMEDIATE CODE: R150

R250 Receipt of annual fees

Free format text: JAPANESE INTERMEDIATE CODE: R250

R250 Receipt of annual fees

Free format text: JAPANESE INTERMEDIATE CODE: R250

R250 Receipt of annual fees

Free format text: JAPANESE INTERMEDIATE CODE: R250

R250 Receipt of annual fees

Free format text: JAPANESE INTERMEDIATE CODE: R250