JP5421842B2 - Impact analysis device, impact analysis method, and program - Google Patents

Description

  The present invention relates to an influence analysis apparatus, an influence analysis method, and a program for estimating a magnitude relationship of influences between users based on behavior histories of a plurality of users.

  In recent years, social services that provide some kind of service by sharing information between users on a network have attracted attention. For example, a social bookmark service, a photo sharing service, a social network service, an online store, and the like. In these services, action histories in which a plurality of users perform actions such as registration and purchase on objects such as web pages, photographs, and products are recorded. For example, in the case of a social bookmark service, a web page corresponds to an object, and bookmark registration for registering the URL of the web page to a server or the like corresponds to an action. Then, an action history of the bookmark registration is stored.

  In social services, users share behavior histories related to the behaviors of other users, so that information regarding future behaviors suitable for the users themselves can be acquired. For example, by using a social bookmark service, a user shares a favorite object (web page) that another user has registered as a bookmark, and by acquiring information on an object that another user has registered as a bookmark, An object suitable for the user can be found.

  In a conventional social service such as this social bookmark service, it is impossible to explicitly know the influence of one user's action on another user's action. In social network services, friendships between users (hereinafter referred to as “user relations”) can be registered, but even a friend may have a strong influence on a user. For example, there are users with weak influence, and the magnitude relationship of influence between users (hereinafter referred to as “influence relation between users”) is not shown.

Until now, the technique which estimates the influence relationship between users from a user's action has been proposed.
For example, an apparatus that predicts the propagation of information on a network in which the relationship between users is known has been proposed (see Non-Patent Document 1). Moreover, when the relationship between users is known, the technique which estimates the influence relationship between users by the information of both the relationship between users and action history is proposed (refer nonpatent literature 2). These techniques estimate the probability of influence on a user's friend based on whether a user's friend later selected a common object.

Saito, K., Nakano, R., and Kimura, M. 2008. Prediction of Information Diffusion Probabilities for Independent Cascade Model.In Proceedings of the 12th international Conference on Knowledge-Based intelligent information and Engineering Systems, Part III (Zagreb, Croatia , September 03-05, 2008) .I. Lovrek, RJ Howlett, and LC Jain, Eds. Lecture Notes In Artificial Intelligence, vol. 5179.Springer-Verlag, Berlin, Heidelberg, 67-75, 2008. Goyal, A., Bonchi, F., and Lakshmanan, LV 2010. Learning influence probabilities in social networks.In Proceedings of the Third ACM international Conference on Web Search and Data Mining (New York, New York, USA, February 04-06 , 2010) WSDM '10. ACM, New York, NY, 241-250.

However, the techniques described in Non-Patent Document 1 and Non-Patent Document 2 have a problem that the influence relationship between users cannot be estimated when the relationship between users (friend relationship) is unknown.
If you can estimate the influence relationship between users, you can clearly indicate the user who has a strong influence relationship to a certain user, predict the user's behavior and specify the object, and predict the popularity of the object Is possible.

  The present invention has been made in view of such a background, and the present invention is an influence analysis device capable of estimating an influence relationship between users even when a friend relationship between users is unknown, It is an object to provide an influence analysis method and program.

  In order to solve the above-mentioned problem, the invention according to claim 1, in which the users share an action history related to selection of another user's object via a communication network, so that the future suitable for the user himself / herself can be obtained. An influence analysis apparatus that is used in a social service that acquires information related to selection of an object, and that analyzes the magnitude relationship of influence between the users who use the social service based on the user's behavior history, the user Each time, the user's action history is acquired as action history information including time information when the object is selected, and the action history information is stored in the storage unit. Using each user to indicate the probability that the object included in the behavior history of the other user is selected by the user. The user object selection probability is calculated for each object, the user object selection probability calculation unit that stores the calculated user object selection probability in the storage unit, the action history information, and the user object selection probability, and the user And an influence probability estimation unit that calculates an influence probability indicating the probability of being influenced by the other user by maximizing a likelihood indicating the likelihood of the estimated value of the influence probability. The influence analysis device was used.

  In addition, according to the sixth aspect of the present invention, information on selection of a future object suitable for the user himself / herself can be obtained by sharing an action history regarding selection of another user's object between users via a communication network. An influence analysis method used in an influence analysis device that is used for an acquired social service and analyzes an influence relationship between the users using the social service based on an action history of the user, The influence analysis apparatus acquires, for each user, the action history of the user as action history information including time information when the selection of the object is performed, and stores the action history information in a storage unit; Using the action history information, for each user, an object included in the action history of the other user is the user. Calculating a user object selection probability indicating the probability of being selected for each object, storing the calculated user object selection probability in the storage unit, using the action history information and the user object selection probability, Calculating an influence probability indicating a probability that the user will be influenced by the other user by maximizing a likelihood indicating a plausibility of the estimated value of the influence probability. Force analysis method.

By doing so, the influence analysis apparatus calculates a user object selection probability, which is a probability that an object included in another user's action history is selected by the user, using the action history information. Then, using the calculated user object selection probability and the action history information, the influence probability that is the probability that a certain user is influenced by another user is maximized and the likelihood indicating the plausibility of the estimated value of the influence probability is maximized. It can be calculated by doing.
Therefore, according to the present invention, it is possible to estimate the influence relationship between users even when the relationship (friend relationship) between users is unknown.

According to the second aspect of the present invention, the influence probability of the other users whose influence probability is higher than a predetermined value is exerted on the user by using the influence probability of each of the other users calculated by the influence probability estimation unit. The influence analyzing apparatus according to claim 1, further comprising a user recommendation unit that extracts as a strong user .

The invention according to claim 7 further executes a step of using the calculated influence probability to extract the other user whose influence probability is higher than a predetermined value as a user having a strong influence on the user. It was influential analysis method according to claim 6, characterized in that.

By doing in this way, other users with high influence probability can be extracted using the influence probability of each other user calculated by the influence analysis apparatus, and can be clearly shown as a user having a strong influence relationship.

The invention according to claim 3 uses the behavior history information to determine, for each user, a probability that an object newly registered as a selection target in the communication network is selected from a predetermined time in the past to the present time. The latest object selection probability is calculated, and the latest object selection probability calculation unit that stores the calculated recent object selection probability in the storage unit, and the action history information is used, for each object, the object is the other object A popular object selection probability calculation that calculates a popular object selection probability indicating a probability that the object is selected by the user in association with the number selected by the user, and stores the calculated popular object selection probability in the storage unit. The object is selected at random using the action history information Random object selection probability calculating unit that calculates a random object selection probability indicating the probability of being selected, and storing the calculated random object selection probability in the storage unit, and using the action history information, any of the users selects A new object selection probability calculation unit that calculates a new object selection probability indicating a probability that the object that is not newly selected is newly selected, and stores the calculated new object selection probability in the storage unit, the influence probability The estimation unit uses the behavior history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability to calculate the influence probability as the likelihood. Calculated by maximizing the degree Was influential analysis apparatus according to claim 1 or claim 2, characterized in.

In the invention according to claim 8, an object newly registered as a selection target in the communication network is selected for each user from the predetermined time in the past to the present time using the action history information. Calculating a recent object selection probability indicating a probability, storing the calculated recent object selection probability in the storage unit, and using the action history information, the object is selected by the other user for each object Calculating a popular object selection probability indicating the probability that the object is selected by the user in association with the calculated number, storing the calculated popular object selection probability in the storage unit, and the action history information. Use random object selection to indicate the probability that the object will be selected at random. Calculating the probability, storing the calculated random object selection probability in the storage unit, and using the action history information, the probability that the object not selected by any of the users is newly selected Calculating the new object selection probability to be stored, and storing the calculated new object selection probability in the storage unit, and calculating the influence probability by maximizing the likelihood. The influence probability is calculated using behavior history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability. The influence analysis method according to item 6 or claim 7 is adopted.

By doing so, the influence analysis apparatus can calculate the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability, and can calculate the influence probability with higher accuracy.

The invention according to claim 4 is the influence probability calculated by the influence probability estimation unit, the action history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, For each object, the selection probability for the user to select the object is calculated for each object, and the object for which the calculated selection probability is higher than a predetermined value is selected as an object for the user to select in the future. It was influential analyzer according to 請 Motomeko 3 you further comprising an object selection prediction unit that extracts.

The invention according to claim 9 is characterized in that the influence probability, the action history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability. For each of the objects, the step of calculating a selection probability that the user selects the object and extracting the object having the calculated selection probability higher than a predetermined value as an object to be selected by the user in the future is further executed. It was influential analysis method according to 請 Motomeko 8 you, characterized in that.

  By doing in this way, the influence analysis apparatus can calculate the selection probability that the user selects each object, and can extract an object with a high selection probability as an object that the user selects in the future.

  The invention according to claim 5 calculates the probability of selecting the object for all the users, using the selection probability for each object calculated by the object selection prediction unit for each user, The influence analyzing apparatus according to claim 4, further comprising a popular object prediction unit that extracts the object having a calculated probability higher than a predetermined value as a popular object.

  The invention according to claim 10 calculates the probability of selecting the object for all of the users using the selection probability of the object calculated for each user, and the calculated probability is The influence analyzing method according to claim 9, further comprising the step of extracting the object higher than a predetermined value as a popular object.

  In this way, the probability of selecting the object can be calculated for all users, and an object with a high probability can be predicted as a popular object in the future.

  The invention according to claim 11 is a program for executing the influence analysis method according to any one of claims 6 to 10 on a general computer.

  According to such a program, the influence analysis method according to claims 6 to 10 can be executed by a general computer.

  According to the present invention, it is possible to provide an influence analysis apparatus, an influence analysis method, and a program capable of estimating an influence relation between users even when a friendship between users is unknown. .

It is a functional block diagram which shows the structural example of the influence analysis apparatus which concerns on this embodiment. It is a sequence diagram which shows the flow of a process of the influence analysis method which concerns on this embodiment. It is a sequence diagram which shows the flow of the production | generation process of the selection prediction information of the object selection prediction part which concerns on this embodiment. It is the figure which compared the correct answer rate about the influence analysis method and collaborative filtering which concern on this embodiment. It is a figure which shows the experimental result which specifies the user with strong influence about the influence analysis method and comparative example which concern on this embodiment.

  Next, modes for carrying out the present invention (referred to as “embodiments”) will be described in detail with reference to the drawings as appropriate.

  In the present embodiment, the social bookmark data in the social bookmark service among the social services will be described. However, if it is the action history information of the user with time, the purchase data, the photo and video sharing site, the data of the product review site Etc. are also applicable.

First, action history information handled in this embodiment will be described.
In the social bookmark service, registering the URL of a web page on a server or the like corresponds to selection of an object. And social bookmark data which is action history information including the time information of the registration time of a web page is set to S = <U, D, Y>. Here, U is a user set, D is a shared object set, and Y is a bookmark history set. The bookmark history (u, d, t) εY indicates that the user uεU bookmarked the object dεD at time t.

  First, consider a case where a user u bookmarks an object, influenced by another user u ′. In other words, when selecting an object to be bookmarked, it is assumed that the user u selects one other user u ′ based on the user-specific probability, and selects from the objects bookmarked by the user u ′. . At this time, for example, the following (formula 1) can be used as the probability that the user u bookmarks the object d at time t.

Here, P (u ′ | u) is a probability that the user u is influenced by another user u ′ with respect to the bookmark (an “impact probability” described later), and Σ _{u ′} P (u ′ | u) = 1. P (u ′ | u) ≧ 0 is satisfied. G (d | u ′, t) is a probability that the object d is included in the bookmark history at time t of the user u ′ (“user object selection probability” described later), and Σ _d G (d | u ′, t) = 1, G (d | u ′, t) ≧ 0 is satisfied. In this case, if P (u ′ | u)> 0, it indicates that the user u ′ affects the user u.

In the present embodiment, the following method is considered as a method for the user to select an object.
(1) Select from the bookmark history of another user u ′. An object selected by this method is called a “user object”.
(2) Select an object newly registered as a selection target recently (from a past predetermined time to the present time). An object selected by this method is called a “recent object”.
(3) Select an object popular with other users u '(the number of bookmarked items is large). An object selected by this method is called a “popular object”.
(4) Select an object at random. An object selected by this method is called a “random object”.
(5) Select an object not selected by any user (not bookmarked). An object selected by this method is called a “new object”.

Here, in order to consider the selection methods (2) to (5) other than the selection method of the object influenced by the other user u ′ shown in (1), in this embodiment, (2), ( As virtual users corresponding to 3), (4), and (5), u _r , u _p , u _s , and u _n are introduced. At this time, it is possible to use the following (Expression 2) instead of (Expression 1) with the probability that the user u bookmarks the object d at time t.

_Here, a _{U A = U∪ {u r,} u p, u s}. If another method is conceivable as a method for selecting an object, another virtual user may be introduced.

<Configuration of influence analysis device>
Next, the configuration of the influence analysis apparatus 1 according to the present embodiment will be specifically described.
FIG. 1 is a functional block diagram illustrating a configuration example of the influence analysis apparatus 1 according to the present embodiment. As shown in FIG. 1, the influence analysis apparatus 1 according to the present embodiment includes an input / output unit 10, a processing unit 20, a memory unit 30, and a storage unit 40.

  Here, the input / output unit 10 is realized by an input / output interface or a communication interface capable of inputting / outputting various information to / from an external device (social bookmark service management server) or the like. The memory unit 30 includes storage means such as a RAM (Random Access Memory), and temporarily stores information necessary for processing by the processing unit 20. The storage unit 40 is a storage means realized by a hard disk or the like, and will be described later, action history information 100, user object selection probability information 201, recent object selection probability information 202, popular object selection probability information 203, random Object selection probability information 204, new object selection probability information 205, and influence probability information 300 are stored. Further, the selection prediction information 250 may be stored, and details will be described later.

  The processing unit 20 is responsible for overall processing of analyzing the influence relationship between users, and includes an input / output information control unit 21, a selection probability calculation unit 22, an influence probability estimation unit 23, a user recommendation unit 24, and an object selection prediction. A unit 25 and a popular object prediction unit 26 are included. The function of the processing unit 20 is realized by, for example, a CPU (Central Processing Unit) developing and executing a program stored in the storage unit 40 of the influence analysis apparatus 1 in the memory unit 30.

  The input / output information control unit 21 acquires the action history information 100 via the input / output unit 10 and stores it in the storage unit 40. Further, the input / output information control unit 21 acquires the influence probability information 300 and the selection prediction information 250 that are the processing results of the influence analysis apparatus 1 and outputs the acquired information through the input / output unit 10.

  The selection probability calculation unit 22 calculates an object selection probability for each of the virtual users described above, a user object selection probability calculation unit 221, a recent object selection probability calculation unit 222, a popular object selection probability calculation unit 223, a random An object selection probability calculation unit 224 and a new object selection probability calculation unit 225 are included.

(User object selection probability calculator)
The user object selection probability calculation unit 221 calculates user object selection probabilities G (d | u ′, t) and dεD for other users u′εU. This will be specifically described below.

First, the user object selection probability calculation unit 221 acquires the input data S _u = <U, D, Y _u >, which is the action history information 100 of the user u, from the storage unit 40. Here, _{Y u} represents the action history information 100 for the user u.
Then, the user object selection probability calculation unit 221 calculates the user object selection probability for each user and each object. This user object selection probability represents which object is easily selected as a bookmark of another user from a user's bookmark history at a certain time. For example, the user object selection probability that the object d is selected from the bookmark history of another user u ′ at time t is calculated by the following (Equation 3).

Here, τ _i represents a target period for calculating the selection probability of the user object.
In addition, H (u ′, t, τ _i ) = | {d | (u ′, d, t ′) ∈Y∧t−t ′ ≦ τ _i } | in the period [t−τ _i , t) Represents the number of objects bookmarked by other users u ′. This target period may be the entire period. In (Expression 3), weighting may be performed so that the recently bookmarked object has a high probability.
After calculating the user object selection probability, the user object selection probability calculation unit 221 stores the calculation result as user object selection probability information 201 in the storage unit 40.

(Recent object selection probability calculator)
The recent object selection probability calculation unit 222 calculates the latest object selection probability G (d | u _r , t), dεD. This will be specifically described below.

First, the recent object selection probability calculation unit 222 acquires the action history information 100 from the storage unit 40.
Then, the nearest object selection probability calculation unit 222 calculates the nearest object selection probability. The recent object selection probability represents how easily a certain object is selected as a user's bookmark from a predetermined past time to the present time (recently). For example, the nearest object selection probability that an object is selected at time t is calculated by the following (formula 4).

Here, τ _r represents a target period in which the latest object selection probability is calculated, and t ′ _d is the time when the object d last appeared.
In addition, M (t, τ _r ) = | {d | (u, d, t ′) ∈Yｔt−t ′ ≦ τ _r } | has an object d in the period [t−τ _r , t). Indicates the number of bookmarks. The target period may be, for example, the most recent day or week. In (Expression 4), the object selection probability is calculated as a uniform distribution recently, but weighting may be performed so that a newer object has a higher probability.
Then, when the recent object selection probability calculation unit 222 calculates the recent object selection probability, the calculation result is stored in the storage unit 40 as the recent object selection probability information 202.

(Popular object selection probability calculator)
Popular object selection probability calculation unit 223, a popular object selection probability _{G (d | u p, t} ), to calculate the d∈D. This will be specifically described below.

First, the popular object selection probability calculation unit 223 acquires the action history information 100 from the storage unit 40.
Then, the popular object selection probability calculation unit 223 calculates a popular object selection probability. This popular object selection probability represents how easily a certain object is selected as a user's bookmark on the premise that a popular (many bookmarked object) is likely to be selected. For example, a popular object selection probability that an object is selected at time t is calculated by the following (formula 5).

Here, τ _p represents a target period for calculating the selection probability of the popular object.
Also, N (d, t, τ _p ) = | {u | (u, d, t ′) ∈Y−t−t ′ ≦ τ _p } | is an object in the period [t−τ _p , t). This represents the number of users who have selected d. The target period may be, for example, the most recent day or week. Further, in (Equation 5), the probability that an object is selected and the number of users are in a relationship represented by a linear function, but may be a relationship such as an exponential distribution.
When the popular object selection probability calculation unit 223 calculates the popular object selection probability, the popular object selection probability calculation unit 223 stores the calculation result as popular object selection probability information 203 in the storage unit 40.

(Random object selection probability calculator)
The random object selection probability calculation unit 224 calculates a random object selection probability G (d | u _s , t), dεD. This will be specifically described below.

First, the random object selection probability calculation unit 224 acquires the action history information 100 from the storage unit 40.
The random object selection probability calculation unit 224 calculates a random object selection probability. For example, the random object selection probability that an object is selected as a bookmark at time t is calculated by the following (formula 6).

Here, M (t, t) represents the number of objects that have appeared (to be selected) until time t.
Then, after calculating the random object selection probability, the random object selection probability calculation unit 224 stores the calculation result as random object selection probability information 204 in the storage unit 40.

(New object selection probability calculator)
The new object selection probability calculation unit 225 calculates a new object selection probability P (u _n | u), uεU for each user. This will be specifically described below.

First, the new object selection probability calculation unit 225 acquires the action history information 100 from the storage unit 40.
Then, the new object selection probability calculation unit 225 calculates a new object selection probability. This new object selection probability is assumed to be the same for any object for a certain user. For example, for a certain user, the new object selection probability is calculated by the following (Equation 7).

Here, l _d is the time when the object d first appears.
When the new object selection probability calculation unit 225 calculates the new object selection probability, the calculation result is stored in the storage unit 40 as new object selection probability information 205.

(Effect probability estimation unit)
Next, the influence probability estimation unit 23 stores the action probability information 100 and the selection probability information calculated by the selection probability calculation unit 22 (user object selection probability information 201, recent object selection probability information 202, Using the popular object selection probability information 203, the random object selection probability information 204, and the new object selection probability information 205), an influence probability that is a probability that the user u is influenced by the user u ′ is calculated.

This influence probability can be calculated, for example, based on a maximum likelihood method that maximizes a likelihood that is a value representing how much the estimated value of the influence probability can explain the input data.
Specifically, the influence probability estimating unit 23 estimates the influence probability by maximizing the likelihood using an EM (Expectation Maximization) algorithm under the following constraint equation (Formula 8).

Since the new object selection probability P (u _n | u) is calculated by the new object selection probability calculation unit 225, the constraint equation is expressed by the following (Equation 9).

Note that the estimation can be performed by using other optimization methods such as Newton's method instead of the EM algorithm.
When using the EM algorithm, the influence probability estimation unit 23 first calculates the posterior probability by the following (Equation 10) in the E (Expectation) step.

  Next, in the M (Maximization) step, the influence probability estimation unit 23 calculates the influence probability by the following (Equation 11).

  Then, the influence probability estimation unit 23 repeats the calculation of the influence probability until the end condition shown in the following (Equation 12) is satisfied, and stores the calculation result as the influence probability information 300 in the storage unit 40.

Here, L _u represents the likelihood of the user u.
As another termination condition, it is also possible to use whether the parameter has converged, repeated a predetermined number of times, whether a predetermined time has elapsed, or the like.

  Further, the influence probability estimation unit 23 may use a posterior probability maximization method, a Markov chain Monte Carlo method, a posterior probability estimation method by variational Bayes, or the like instead of the maximum likelihood method.

  Next, the user recommendation unit 24, the object selection prediction unit 25, and the popular object prediction unit 26 will be described.

  The user recommendation unit 24 obtains the influence probability information 300 calculated by the influence probability estimation unit 23 and stored in the storage unit 40, and ranks other users u ′ related to a certain user u in descending order of influence probability. By outputting the user u ′ having the highest influence probability via the input / output unit 10, the user u ′ having the strongest influence relationship can be clearly indicated for the user u.

The object selection prediction unit 25 includes the influence probability information 300, the action history information 100, the user object selection probability information 201, the recent object selection probability information 202, the popular object selection probability information 203, and the random object selection probability stored in the storage unit 40. Information 204 and new object selection probability information 205 are acquired, and a certain user u predicts which object will be selected in the future.
The object selection prediction unit 25 calculates the probability that the user u selects the object d at time t (object selection probability) using the following (Equation 13).

  The object selection prediction unit 25 ranks the selection probabilities of the objects calculated in (Equation 13) in descending order. Then, the object selection prediction unit 25 stores an object having a selection probability equal to or higher than a predetermined value or a predetermined number of objects from the top in the storage unit 40 as selection prediction information 250. Then, the input / output information control unit 21 outputs the selection prediction information 250, whereby an object with a high probability of being selected by the user u in the future can be specified.

  The popular object prediction unit 26 obtains the selection prediction information 250 calculated by the object selection prediction unit 25 and stored in the storage unit 40, and selects from each user in the future based on the selection probability for each object for all users. An object having a high probability of being played can be calculated, and the object can be predicted as a highly popular object.

<Operation procedure>
Next, the process of the influence analysis method according to the present embodiment will be described along FIG. 2 with reference to FIG.
FIG. 2 is a sequence diagram showing a processing flow of the influence analysis method according to the present embodiment.

  First, the input / output information control unit 21 of the influence analysis apparatus 1 acquires the action history information 100 via the input / output unit 10 and stores it in the storage unit 40 (step S101).

Next, the selection probability calculation unit 22 initializes the parameters so that the total influence probability is Σ _{u ′} P (u ′ | u) = 1 in (Equation 2) described above (step S102). .

  Next, the user object selection probability calculation unit 221 acquires the action history information 100 from the storage unit 40, and calculates the user object selection probability G (d | u ′, t) using (Equation 3) described above. (Step S103). Then, the user object selection probability calculation unit 221 stores the calculation result as user object selection probability information 201 in the storage unit 40.

Subsequently, the recent object selection probability calculation unit 222 acquires the action history information 100 from the storage unit 40, and calculates the latest object selection probability G (d | u _r , t) using (Equation 4) described above. (Step S104). Then, the recent object selection probability calculation unit 222 stores the calculation result in the storage unit 40 as the latest object selection probability information 202.

The popular object selection probability calculation unit 223 acquires the action history information 100 from the storage unit 40, using the above-described (Equation 5), popular object selection probability G | calculate the (d _{u p,} t) ( Step S105). Then, the popular object selection probability calculation unit 223 stores the calculation result in the storage unit 40 as popular object selection probability information 203.

Next, the random object selection probability calculation unit 224 acquires the action history information 100 from the storage unit 40, and calculates the random object selection probability G (d | u _s , t) using (Equation 6) described above. (Step S106). Then, the random object selection probability calculation unit 224 stores the calculation result as random object selection probability information 204 in the storage unit 40.

  Subsequently, the new object selection probability calculation unit 225 selects the first user (X = 1) from all the users (step S107).

Then, the new object selection probability calculation unit 225 calculates the new object selection probability P (u _n | u _x ) using (Equation 7) described above (step S108).

  Next, the influence probability estimation unit 23 calculates the influence probability using the above-described (Expression 10) and (Expression 11) (Step S109).

  Subsequently, the influence probability estimation unit 23 determines whether or not the parameter is converged by satisfying the above-described termination condition of (Equation 12) (step S110).

  If (Equation 12) does not converge (step S110 → No), the influence probability estimation unit 23 returns to step S109 and continues the process, assuming that the end condition is not satisfied. On the other hand, when (Equation 12) converges (step S110 → Yes), the influence probability estimation unit 23 proceeds to the next step S111, assuming that the end condition is satisfied.

  In step S111, the influence probability estimation unit 23 determines whether or not X = the number of users, that is, whether or not all users have been processed. Here, when there is a user who has not yet performed the process (step S111 → No), the next user who has not yet performed the process is selected (step S112), and the process returns to step S108. On the other hand, if X = the number of users, that is, if all users have been processed (step S111 → Yes), the process proceeds to the next step S113.

  In step S113, the influence probability estimation unit 23 stores the calculated influence probability of each user in the storage unit 40 as the influence probability information 300.

In this way, according to the influence analysis device, the influence analysis method, and the program according to the present embodiment, the probability that a certain user is influenced by other users using the action history information 100 of each user. Can be calculated by maximizing the likelihood indicating the plausibility of the estimated value of the impact probability.
Therefore, according to the present invention, it is possible to estimate the influence relationship between users even when the relationship (friend relationship) between users is unknown.

  Next, processing of the user recommendation unit 24, the object selection prediction unit 25, and the popular object prediction unit 26 will be described.

  The user recommendation unit 24 acquires the influence probability information 300 stored in the storage unit 40, ranks the other users u ′ related to the target user u in descending order of the influence probability, and outputs the upper user u ′. To do. Thereby, the user recommendation part 24 can specify a user with a strong influence relation to a user.

  The object selection prediction unit 25 predicts which object a user u will select in the future. FIG. 3 is a sequence diagram showing a flow of processing for generating selection prediction information 250 of the object selection prediction unit 25 according to the present embodiment.

  As shown in FIG. 3, first, the object selection prediction unit 25 receives the action history information 100 and each selection probability information (user object selection probability information 201, recent object selection probability information 202, popular object selection probability from the storage unit 40. Information 203, random object selection probability information 204, new object selection probability information 205) and influence probability information 300 are acquired (step S201).

  Next, the object selection prediction unit 25 calculates an object selection probability, which is a probability that the user u selects the object d at time t, using (Equation 13) described above (step S202).

  Subsequently, the object selection prediction unit 25 executes ranking processing in descending order of selection probability for all objects (step S203).

  Then, the object selection prediction unit 25 extracts the ranking higher than the predetermined value, generates selection prediction information 250, and stores it in the storage unit 40 (step S204).

  By doing in this way, the influence analysis apparatus 1 can calculate the selection probability that a user selects each object, and can extract an object with the high selection probability as an object which the user selects in the future.

  Further, the popular object prediction unit 26 acquires the selection prediction information 250 stored in the storage unit 40, and based on the selection probability for each object for all users, an object that is likely to be selected from each user in the future. And the object can be predicted as a popular object.

<Experimental example of this embodiment>
In order to evaluate the influence analysis apparatus 1 according to the present embodiment, an experiment was performed using action history information 100 collected from a social bookmark site. In this experiment, the influence was analyzed for each set of web pages (objects) to which 50 kinds of popular tags were attached. Here, in order to analyze long-term and short-term influence, the target period for calculating the user object selection probability is τ _i = 30, 60, 90 (days). Further, the target period for calculating the recent object selection probability is τ _r = 1 (day), and the target period for calculating the popular object selection probability is τ _p = 30 (day).

  The correct answer rate was used as an evaluation index for this experiment. The correct answer rate is calculated by the following (formula 14).

は、ユーザｕが時刻ｔ_ｘ以降にブックマークしたｎ個のオブジェクト集合を表す。 Here, top _m (u, t _x ) represents a set of m objects at the top t of user u at time t _x .

Represents the n-number of objects set by the user u has bookmark after time t _x.

  And collaborative filtering was used as a comparison method. This collaborative filtering is a technique that is influenced by users who have similar histories (high cosine similarity).

FIG. 4 is a diagram comparing the correct answer rate between the influence analysis method according to the present embodiment and collaborative filtering. As shown in FIG. 4, the influence analysis method according to the present embodiment has a higher correct answer rate than collaborative filtering. In the collaborative filtering, the correct answer rate decreases as the learning period (action history information acquisition period) becomes longer. This indicates that processing in consideration of the target period (τ _i , τ _r , τ _p ) and the like is important as in the influence analysis method according to the present embodiment.

  Next, it was evaluated whether a user having a large influence could be specified by the influence analysis method according to the present embodiment. This evaluation is performed by a score indicating how much an object bookmarked by a certain user is bookmarked by another user who has a strong influence on the user.

Here, the user columned _{I u} to the user order influence strongly the user u, rank _u 'other user u in _{I u} (u)' as ranking is calculated following Equation (15).

  Then, the score at time t is calculated by the following (Equation 16).

FIG. 5 is a diagram showing an experimental result for identifying a user who has a strong influence. As a comparison method, we used collaborative filtering and random ranking. In FIG. 5, the influence is estimated by setting 500 days as the learning period of the action history, and further increasing the learning period by 50 days to 500 days. Then, for each learning period, the next 7 days were used as test data, and τ _i = 60 (days).
As shown in FIG. 5, the influence analysis method according to the present embodiment can obtain a higher score than the collaborative filtering (ranking users according to similarity) that is a comparison method and the method of ranking randomly. In addition, it was shown that as the learning period becomes longer, it is possible to identify a user having higher influence with higher accuracy.

DESCRIPTION OF SYMBOLS 1 Influence analysis apparatus 10 Input / output part 20 Processing part 21 Input / output information control part 22 Selection probability calculation part 23 Influence probability estimation part 24 User recommendation part 25 Object selection prediction part 26 Popular object prediction part 30 Memory part 40 Storage part 100 Action History information 201 User object selection probability information 202 Recent object selection probability information 203 Popular object selection probability information 204 Random object selection probability information 205 New object selection probability information 221 User object selection probability calculation unit 222 Recent object selection probability calculation unit 223 Popular object selection Probability calculator 224 Random object selection probability calculator 225 New object selection probability calculator 250 Selection prediction information 300 Influence probability information

Claims (11)

The social service is used for a social service that acquires information on selection of a future object suitable for the user by sharing an action history regarding selection of another user's object with each other via a communication network. An influence analysis device that analyzes the magnitude relationship of influence between the users using the user based on the action history of the user,
For each user, an input / output information control unit that acquires the behavior history of the user as behavior history information including time information when the selection of the object is performed, and stores the behavior history information in a storage unit;
Using the action history information, for each user, a user object selection probability indicating the probability that an object included in the action history of the other user is selected by the user is calculated for each object, and the calculated user A user object selection probability calculation unit for storing an object selection probability in the storage unit;
Using the action history information and the user object selection probability, maximizing an influence probability indicating the probability that the user is affected by the other user and indicating a likelihood of the estimated value of the influence probability An impact probability estimator calculated by
An influence analysis apparatus comprising: A user recommendation unit that extracts the other users whose influence probability is higher than a predetermined value as a user having a strong influence on the user by using the influence probability of each of the other users calculated by the influence probability estimation unit.
The influence analyzing apparatus according to claim 1, further comprising: Using the action history information, for each user, calculating a recent object selection probability indicating a probability that an object newly registered as a selection target in the communication network is selected from a predetermined time in the past to the current time, A nearest object selection probability calculation unit for storing the calculated nearest object selection probability in the storage unit;
Using the action history information, for each object, calculate the popular object selection probability indicating the probability that the object is selected by the user, in association with the number of the object selected by the other user, A popular object selection probability calculation unit that stores the calculated popular object selection probability in the storage unit;
A random object selection probability calculation unit that calculates a random object selection probability indicating a probability that the object is randomly selected using the action history information, and stores the calculated random object selection probability in the storage unit;
Using the action history information, calculate a new object selection probability indicating a probability that the object not selected by any of the users is newly selected, and store the calculated new object selection probability in the storage unit A new object selection probability calculator to
Further comprising
The influence probability estimation unit
Using the action history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability, the influence probability and the likelihood are maximized. To calculate by doing
Impact analysis apparatus according to claim 1 or claim 2, characterized in. Using the influence probability, the action history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability calculated by the influence probability estimation unit And an object selection prediction unit that calculates a selection probability for the user to select the object for each object, and extracts the object having the calculated selection probability higher than a predetermined value as an object to be selected by the user in the future. impact analysis apparatus according to 請 Motomeko 3 you wherein a. Using the selection probability for each object calculated by the object selection prediction unit for each user, the probability of selecting the object is calculated for all the users, and the calculated probability is higher than a predetermined value. The influence analysis apparatus according to claim 4, further comprising a popular object prediction unit that extracts an object as a popular object. The social service is used for a social service that acquires information on selection of a future object suitable for the user by sharing an action history regarding selection of another user's object with each other via a communication network. An influence analysis method used in an influence analysis apparatus for analyzing the magnitude relationship of influence between the users using the user based on the user's behavior history,
The influence analysis device
For each user, obtaining the user's action history as action history information including time information for executing the selection of the object, and storing the action history information in a storage unit;
Using the action history information, for each user, a user object selection probability indicating the probability that an object included in the action history of the other user is selected by the user is calculated for each object, and the calculated user Storing an object selection probability in the storage unit;
Using the action history information and the user object selection probability, maximizing an influence probability indicating the probability that the user is affected by the other user and indicating a likelihood of the estimated value of the influence probability The step of calculating by
The influence analysis method characterized by performing. Using the calculated influence probability, extracting the other user whose influence probability is higher than a predetermined value as a user having strong influence on the user;
The influence analysis method according to claim 6, further executed . Using the action history information, for each user, calculating a recent object selection probability indicating a probability that an object newly registered as a selection target in the communication network is selected from a predetermined time in the past to the current time, Storing the calculated recent object selection probability in the storage unit;
Using the action history information, for each object, calculate the popular object selection probability indicating the probability that the object is selected by the user, in association with the number of the object selected by the other user, Storing the calculated popularity object selection probability in the storage unit;
Using the action history information, calculating a random object selection probability indicating a probability that the object is randomly selected, and storing the calculated random object selection probability in the storage unit;
Using the action history information, calculate a new object selection probability indicating a probability that the object not selected by any of the users is newly selected, and store the calculated new object selection probability in the storage unit And steps to
And run further
In the step of calculating the influence probability by maximizing the likelihood, the action history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new Using the object selection probability to calculate the influence probability
Impact analysis method according to claim 6 or claim 7, characterized in. Using the influence probability, the behavior history information, the user object selection probability, the recent object selection probability, the popular object selection probability, the random object selection probability, and the new object selection probability, for each object, the user There calculate the selection probability of selecting the object, the calculated selection probability is you, characterized in that it further performs the step of extracting the higher the object than a predetermined value as an object which the user selects a future 請 Motomeko 8 The impact analysis method described in 1. Using the selection probability for each object calculated for each user, the probability of selecting the object is calculated for all the users, and the object having the calculated probability higher than a predetermined value is highly popular. The influence analyzing method according to claim 9, further comprising the step of extracting as an object.   A program for causing a computer to execute the influence analysis method according to any one of claims 6 to 10.

Images