CN108052961B - Multi-factor decision-making method for activity recommendation of active social network users - Google Patents

Abstract

The invention relates to a multi-factor decision-making method for activity recommendation of active social network users, which comprises the following steps: an activity-host-user ESU graph model is constructed facing an activity social network EBSN, and the social influence of the activity on the user in the graph is calculated; calculating the correlation between the activity and the content of the activity in which the user participates; calculating the relevance of the activity and the activity place in which the user participates; calculating the time correlation between the activity and the activity in which the user participates; based on the above four factors, a classical J48 decision tree algorithm is used to decide whether a user will participate in an activity. The method of the invention predicts whether the user participates in the activity or not by utilizing a plurality of factors of social influence, content relevance, location relevance and time relevance of the activity and the user, reasonably utilizes the characteristics of the activity social network EBSN, and is suitable for activity recommendation of the activity social network EBSN.

Description

Multi-factor decision-making method for activity recommendation of activity social network users

Technical Field

The invention relates to an information mining technology, in particular to a multi-factor decision-making method for recommending activities of active social network users.

Background

In recent years, active-based Social Networks (EBSNs) have attracted attention from researchers, mainly because they provide a platform for users to organize, participate in, and share Social activities in an online manner, and users can participate in real offline activities, such as Meetup, Plancast, Facebook Event, and the same city as bean. In the face of many activities that are held at different times and locations on EBSNs, it takes a significant amount of time for a user to find an activity of interest. On EBSNs, activities are automatically and accurately recommended to users, and therefore the method has important practical significance in enriching amateur life, expanding social relations and enjoying team entertainment for the users.

Unlike previous social networks, EBSNs have their unique features: (1) most of the events are held by a host, the held events comprise information such as event types, event contents, event time, event places, people interested in/participating in the events and the like, and the connotation of the events is richer; (2) a large number of entities, including activities, users, sponsors and the like, are arranged on the EBSNs, and unique complex social relationships are constructed among the entities; (3) the participation of users in an activity is not only attracted by the content of the activity, but also influenced by social relationships, the time and place of the activity, and the like.

Conference discourse set published in us 2012: the 18th Knowledge Discovery and Data Mining conference in 2012 (18th ACM SIGKDD conference on Knowledge Discovery and Data Mining), titled: active social networking: the online and offline social networks (events-based social networks: linking the online and offline social networks) were connected by Liu X, He Q, tie Y, Lee WC, McPherson J, Han J, which for the first time defined the concept of active social networks EBSNs, which are believed to connect the online and online social networks, as a new type of social media.

Conference discourse published in the us in 2015: the 2015 Conference on general Computing (ACM International Joint Conference on Pervasive and Ubiquitous Computing) is titled: will i invite who will participate in the meeting? The system integrates user preference and social activity with maximized influence (the Who has I in the service for my party is the combining user preference and the influece maximum knowledge for social events), and the authors are Yu Z, Du R, Guo B, Xu H, Gu T, Wang Z and Zhang D.

Journal published in the netherlands in 2017: neural computing (neuronoputing), entitled: the active social network is used for a mixed collaborative filtering model (A hybrid collaborative filtering model for social influence prediction in event-based social networks), the author is Li X, Cheng X, Su S, Li SC, Yang JY, the document provides a mixed collaborative filtering model MF-EUN which fuses the neighbors of the activity and the user, and provides a neighbor discovery method based on additional information in order to solve the sparsity of a social influence matrix. It is not accurate enough to model EBSNs between users, between users and campaigns, and between campaigns and sponsors, and also does not take into account the impact of time elements on campaign recommendations.

The inventor finds that in the aspect of activity recommendation of active social network users, the modeling of the active social network is not accurate enough, and the feature analysis and utilization of the active social network are not comprehensive enough.

Disclosure of Invention

In view of the problems and the defects of the prior art, the technical problem to be solved by the invention is to provide the multi-factor decision method for recommending the activity of the activity social network user, the method models the activity social network more reasonably, comprehensively utilizes the characteristics of the activity social network, and is more suitable for recommending the activity of the activity social network user.

In order to solve the problems, the invention adopts the following technical scheme:

the invention discloses a multi-factor decision-making method for a user to participate in activity recommendation in an activity social network, which comprises the following specific steps:

A. activity-oriented social network EBSN builds an activity-host-user ESU graph model, and calculates activity e in the graph_iFor user u_jResulting social influence si (e)_i，u_j)；

B. Computing activity e_iAnd user u_jAttended campaign content relevance c (e)_i，u_j)；

C. Computing activity e_iAnd user u_jAttended event location relevance l (e)_i，u_j)；

D. Computing activity e_iAnd user u_jTemporal relevance of attended events t (e)_i，u_j)；

E. Based on the four factors, the classical J48 decision tree algorithm is used for deciding the user u_jWhether or not to participate in the event e_i。

In the multi-factor decision method for recommending activities participated by users of the activity social network, the EBSN oriented to the activity social network in the step (A) constructs an activity-host-user ESU graph model, and calculates the activity e in the graph_iFor user u_jResulting social influence si (e)_i，u_j) The preferable specific steps are as follows:

a1, constructing an activity-host-user ESU graph model, wherein the ESU graph model is described as a binary group: (Entity, relationship), wherein Entity represents an Entity, including an activity set E, a host set S, and a user set U; relationship of relationship representation, including host-activity relationship set SE, host-user relationship set SU, user-activity relationship set UE and relationship set UU between users;

a2, calculating host-activity relation weight, let UN(s)_j) As a host s_jNumber of users involved, let UN (e)_i) Is as s_jMiddle pair of activities e_iNumber of interested/participating users, host s_j-activity e_iThe relationship weight of (1) is:

a3, calculating the weight of host-user relationship, and letting SN (u)_i) For user u_iThe number of hosts added, EN(s)_j) As a host s_jThe number of events ever held, SEN (u)_i) For user u_iHosts s participating once_jThe number of events, the host s_j-user u_iThe relationship weight of (1) is:

a4, calculating the weight of the user-activity relationship, let E (u)_j) For user u_jSet of activities, s, participated in_jIs an activity e_iThe host, ES (u)_j) For user u_jTake part in_jThe set of held events, then user u_j-an activity e_iThe relationship weight of (1) is:

a5, calculating the weight of user-user relationship, and making E (u) be the activity set of user uE (v) an activity set joined by user v, s (u) a host set joined by user u, and s (v) a host set joined by user v, the relationship weight of user u-user v is:

a6, and iterative calculation of influence of the host, the user and the activity in the ESU diagram are respectively shown as formulas (1-1), (1-2) and (1-3):

S_k＝c(W_SEE_k-1+W_SUU_k-1)+(1-c)se_j (1-1)

U_k＝c(W_SUS_k-1+W_UUU_k-1)+(1-c)ue_j (1-2)

E_k＝c(W_SES_k-1+W_UEU_k-1)+(1-c)ee_j (1-3)

se in formulae (1-1), (1-2), (1-3)_j、ue_jAnd ee_jRespectively, the jth dimension is a unit vector of 1, and 1-c are return nodes se_j、ue_jAnd ee_jProbability of (W)_SE、W_SU、W_UUAnd W_UEHost-event, host-user, user-user, and user-event relationship weights, respectively.

In the multi-factor decision method for recommending activities of users in active social network, the calculation activity e in the step (B)_iAnd user u_jAttended campaign content relevance c (e)_i，u_j) The preferable specific steps are as follows:

b1, user u_jThe set of activities that have been attended is denoted E (u)_j) Computing similarity of activities using type labels of activities, CF (e)_i) Representing an activity e_iType tag set of CF (E (u)_j) Represents user u)_jA set of type tags that have participated in the activity;

B2、activity e_iAnd E (u)_j) The similarity calculation method comprises

In the multi-factor decision method for recommending activities of users in an active social network, the calculation activity e in the step (C)_iAnd user u_jAttended event location relevance l (e)_i，u_j) The preferable specific steps are as follows:

C1、

representing an activity e_iThe location of (a) is determined,

representing user u_jThe location of (a) is determined,

to represent

And

the distance of (d);

c2, Activity e_iLocation and user u_jThe method for calculating the relevance of the place comprises the following steps:

in the multi-factor decision method for recommending activities of users in an active social network, the calculation and calculation activity e in the step (D)_iAnd user u_jTemporal relevance of attended events t (e)_i，u_j) The preferable specific steps are as follows:

d1, let the user have beenParticipated and Activity e_iThe same type of activity set is denoted as E (u)_j) (each activity has a type tag), Activity e_iThe holding time is recorded as

Same type of activity e that the user has participated in_k′∈E(u_j) Is recorded as

Then activity e_iTime and type of activity e_kThe time offset of' is:

d2, Activity e_iWith user u_jThe time deviation calculating method comprises the following steps:

min represents taking the minimum value.

Compared with the prior art, the multi-factor decision method for recommending the activity of the active social network user, disclosed by the invention, more reasonably models the active social network, comprehensively utilizes the characteristics of the active social network, and is more suitable for recommending the activity of the active social network user.

Drawings

FIG. 1 is a flow diagram of a multi-factor decision-making method for an active social network user to engage in an activity recommendation of the present invention;

FIG. 2 is a diagram of the activity-host-user ESU graph model constructed by the activity-oriented social network EBSN described in step 101 of FIG. 1, where activity e is calculated_iFor user u_jResulting social influence s_i(e_i，u_j) And (4) a flow chart.

Detailed Description

The following describes the implementation of the present invention in further detail with reference to the accompanying drawings and the detailed description.

Referring to fig. 1, the multi-factor decision method for a user to participate in activity recommendation in an active social network of the present invention comprises the following steps:

step 101, constructing an activity-host-user ESU graph model facing to an activity social network EBSN, and calculating an activity e in the graph_iFor user u_jResulting social influence si (e)_i，u_j) Referring to fig. 2, the specific steps are as follows:

step 201, constructing an activity-host-user ESU graph model, wherein the ESU graph model is described as a binary group: (Entity, relationship), wherein Entity represents an Entity, including an activity set E, a host set S, and a user set U; relationship of relationship representation, including host-activity relationship set SE, host-user relationship set SU, user-activity relationship set UE and relationship set UU between users;

step 202, let UN(s)_j) As host s_jNumber of users involved, let UN (e)_i) Is as s_jMiddle pair of activities e_iNumber of users interested/participating. Then the host s_j-an activity e_iThe relationship weight of (1) is:

equation (1) makes use of the host s_jMiddle participation activity e_iNumber of users, and s_jThe number of all users in (a) measures the host's relationship weight to the event. se(s)_j，e_i) The larger the value, the host s is indicated_jThe more users in s_jThe event held.

Step 203, let SN (u)_i) For user u_iThe number of hosts added, EN(s)_j) As a host s_jThe number of events ever held, SEN (u)_i) For user u_iHosts s participating once_jThe number of events, the host s_j-user u_iThe relationship weight of (1) is:

equation (2) in the first half illustrates user u_iParticipant s_jThe more activities, the host s_jFor user u_iThe greater the effect of (c); the latter half of equation (2) takes into account user u_iThe number of other sponsors added is intuitive, and the more sponsors a user participates in, the more sponsors s_jThe less the impact on the user.

Step 204, let E (u)_j) For user u_jSet of activities, s, participated in_jIs an activity e_iHost, ES (u)_j) For user u_jTake part in_jA set of hosted activities. Then user u_j-an activity e_iThe relationship weight of (1) is:

equation (3) illustrates user u_jParticipated in activity e_iThe more events the host has, user u_jFor activity e_iThe greater the effect of (c).

Step 205, let e (u) be the activity set attended by user u, e (v) be the activity set attended by user v, s (u) be the host set attended by user u, s (v) be the host set attended by user v, and then the relationship weight between user u and user v is:

equation (4) comprehensively considers the situation of the co-event in which the user u and the user v participate and the co-sponsor joined, and α and β are parameters for balancing the weight between the co-event and the co-sponsor.

It should be noted that, in turn, user v may weigh user u differently, since v and u may be interested in a different number of hosts and may be participating in a different number of activities, but the calculation method is similar to equation (4).

Step 206, the ESU graph is a directed graph, and the importance of the three types of entities is reflected differently. Host sides_iIs dependent on s_iAll events held and all users involved have an influence on it, so s_iThe calculation of (b) needs to consider its out-degree (Hubs), and PageRank considers the in-degree of the page. The importance of an active e-node depends on the host and the impact that the user has on it. The importance of user u node depends on the host and the impact it has on his users in mind. Therefore, the activity of the ESU graph and the calculation of the importance of the user should take into account their degree of attendance (authorizations).

Therefore, when calculating the importance of each type of node of the ESU, iterative calculation needs to be performed from the perspective of both out-degree and in-degree. For this reason, we propose a node importance calculation method based on a bidirectional restart random walk algorithm (BD-RWR) on an ESU graph. The importance of the event host, the importance of the user and the event are iteratively calculated as shown in formulas (5), (6) and (7), respectively:

S_k＝c(W_SEE_k-1+W_SUU_k-1)+(1-c)se_j (5)

U_k＝c(W_SUS_k-1+W_UUU_k-1)+(1-c)ue_j (6)

E_k＝c(W_SES_k-1+W_UEU_k-1)+(1-c)ee_j (7)

se in formulae (5), (6) and (7)_j、ue_jAnd ee_jRespectively, unit vectors of which the j-th dimension is 1. 1-c being a return node se_j、ue_jAnd ee_jThe probability of (c). W_SE、W_SU、W_UUAnd W_UEHost-event, host-user, user-user, and user-event relationship weights, respectively.

After a plurality of iterations, each user u is obtained_iThe greater the node importance, the more activity e is illustrated_iThe greater the social impact on the ESU graph on the user, the more likely the user is to participate in activity e_i。

Step 102, calculating Activity e_iAnd user u_jAttended campaign content relevance c (e)_i，u_j)：

Given a particular activity e_iLater, when recommending to the user, e needs to be considered_iActive content and user u_jSimilarity of activities that have been engaged. Suppose user u_jThe set of activities that have been attended is denoted E (u)_j)。

Existing methods in computing Activity e_iActivity E (u) with the user_j) When the similarity is measured, most subjects are extracted from the activities by using an LDA model, and then the similarity of the activities is measured by using cosine similarity.

However, the EBSNs activity data is often concise, so that the sparsity of text data is increased, and the topics are not easy to be accurately extracted from the text. And all activities are mixed together to extract a theme, so that the activities with small participation amount of users are easily submerged.

Computing similarity of activities using type labels of activities, Activity e_iAnd E (u)_j) The similarity calculation method of (2) is shown in equation (8):

in formula (8), CF (e)_i) Representing an activity e_iType tag set of CF (E (u)_j) Represents user u)_jA set of type tags that participated in the activity.

Step 103, calculating activity e_iAnd user u_jAttended event location relevance l (e)_i，u_j)：

The event location refers to the real location held under the event line. Unlike traditional social media, the hosting of offline activities is a unique feature of EBSNs. Thus, the location factor has an important role in deciding whether to recommend an activity to the user.

Activity e_iLocation and user u_jThe correlation calculation of the location is shown in equation (9):

in the formula (9), the reaction mixture is,

representing an activity e_iThe location of (a) is determined,

representing user u_jThe location of (a) is determined,

to represent

And

the distance of (c). The places have a longitude and latitude value, and the distance between the two places is easy to calculate.

Step 104, calculating Activity e_iAnd user u_jTemporal relevance of attended events t (e)_i，u_j)；

People engaging in activities are limited by time factors, typically manifested as a periodic nature of days and weeks. For example, a habit is to attend an activity every day after work, or every weekend. Thus, the time factor has a significant impact on whether a user can participate in an activity.

Associating the user with the activity e_iThe same type of activity set is denoted as E (u)_j) (each activity has a type tag), Activity e_iThe holding time is recorded as

Same type of activity e that the user has participated in_k′∈E(u_j) The holding time is recorded as

Then activity e_iTime and type of activity e_kThe time offset of' is:

activity e_iWith user u_jThe time deviation calculation method (2) is shown in equation (11):

equation (11) in calculating Activity e_iWith user u_jWhen the time deviation is obtained, on one hand, the type of the activity is considered, and because of different time, the activities of the users are different, such as meeting at night and traveling on weekends; on the other hand, use user u_jIn the event of participation_iThe deviation of the activity time with the closest holding time is used as a measurement standard, so that the deviation caused by averaging can be avoided.

Step 105, on the basis of the four factors, determining a user u by using a classic J48 decision tree algorithm_jWhether or not to participate in the event e_i。

And comparing the effectiveness of the users participating in the activity by using four different activity social network user participation activity recommendation methods. The four methods are as follows:

● basic matrix decomposition method, which only considers the user-activity join result, the weight of the elements in the matrix is calculated by formula (3) in this specification, abbreviated as MF;

●, matrix decomposition related trust propagation is used, user-activity participation results and simple social relations are integrated, and the construction of the user social relations uses the formula (4) of the specification, which is abbreviated as SocialMF and parameter setting;

● hybrid collaborative Filtering model, abbreviated MF-EUN-ER, with threshold of influence set to 0.5, active neighbor k₁50, neighbor k of user₂60, 20 area clustering numbers t;

● since decision tree method is easy to read and helpful for manual analysis, the effect of 4 factors mentioned herein is examined by using the classic J48 decision tree method, abbreviated as F4-DT-ER;

the invention selects the data of the same city of bean cotyledon as the experimental corpus. The bean has more than 3 million users in the same city, and the activity of the offline activities is high.

To reduce the workload of data analysis, and at the same time to avoid the impact of some inactive users or activities on the overall experimental results, some of the very inactive users and activities are deleted. Less than 5 users interested/engaged in the activity are considered cold start users, accounting for 9% of the total number of all users; if a certain activity is interested/attended by less than 8 people, it is considered as a cold start activity, accounting for 6% of the total number of all activities. After 2 months of data had been pre-processed, a total of 8663 users, 17822 campaigns, and 294 sponsors were obtained.

In EBSNs, a user indicates whether an activity is attended or not attended, and the activity attendance question can be converted into a binary question. During the experiment, 80% of the data sets were randomly selected for training and the other 20% were selected for testing.

Using F1-measure ═ (2 × P × R)/(P + R) as an evaluation index, where P is the accuracy of the recommended user participating in the activity and R is the recall of the recommended user participating in the activity.

The results obtained by the four methods are shown in table 1 for the recommended overall results of the user's activities.

TABLE 1 Overall recommendation results obtained by the four methods

As can be seen from Table 1, the method F4-DT-ER provided by the invention has the most ideal result, F1-measure is 0.86, and then MF-EUN-ER method is carried out, and F1-measure is 0.80. Of the four methods, MF has the worst effect, F1-measure is 0.67, the main reason is that it only starts from the relationship between the user and the activity, and omits many other relationships on the EBSNs, similar to the conventional user-item recommendation method, which also indicates that the user participation activity recommendation effect of the conventional recommendation method for EBSNs is relatively poor. The effect of the method, namely, the social MF, is improved, and the F1-measure is 0.72, because the social relationship and the trust propagation among the users are considered, but the social relationship among the users is only considered, and the utilization of the relationship among other entities is insufficient. According to the characteristics of EBSNs, the method MF-EUN-ER takes the activity content, the activity site and the social relationship into consideration, the obtained result is improved to a great extent, and the F1-measure value is improved by 0.13 compared with NMF and is improved by 0.08 compared with SocialMF. The F4-DT-ER deep analysis provided by the invention utilizes the unique social relationship characteristics of EBSNs, provides a user influence calculation method of an ESU diagram, well utilizes the content, the location and the time factor of the activity, and obtains the best effect.

This shows that, for EBSNs, it is necessary to grasp their unique features and comprehensively consider multiple factors to obtain a better result of recommending the user participation in the activity. The ESU graph representation model, the social influence calculation method and the activity recommendation multi-factor decision model provided by the invention are unique characteristics of the EBSNs.

Claims (4)

1. A multi-factor decision-making method for recommending activities of users in an active social network comprises the following specific steps:

A. activity-oriented social network EBSN builds an activity-host-user ESU graph model, and calculates activity e in the graph_iFor user u_jResulting social influence si (e)_i,u_j)；

B. Computing activity e_iAnd user u_jAttended campaign content relevance c (e)_i,u_j)；

C. Computing activity e_iAnd user u_jAttended event location relevance l (e)_i,u_j)；

D. Computing activity e_iAnd user u_jTemporal relevance of attended events t (e)_i,u_j)；

E. In the above social influence si (e)_i,u_j) Activity content relevance c (e)_i,u_j) Activity site relevance l (e)_i,u_j) Activity time correlation t (e)_i,u_j) Based on four factors, the classical J48 decision tree algorithm is used for deciding the user u_jWhether or not to participate in the event e_i；

Constructing an activity-host-user ESU graph model by the activity-oriented social network EBSN in the step (A), and calculating an activity e in the graph_iFor user u_jResulting social influence si (e)_i,u_j) The method comprises the following specific steps:

a1, constructing an activity-host-user ESU graph model, wherein the ESU graph model is described as a binary group: (Entity, relationship), wherein Entity represents an Entity, including an activity set E, a host set S, and a user set U; relationship of relationship representation, including host-activity relationship set SE, host-user relationship set SU, user-activity relationship set UE and relationship set UU between users;

a2, calculating host-activity relation weight, let UN(s)_j) As a host s_jNumber of users involved, let UN (e)_i) Is s is_jMiddle pair of activities e_iThe number of interested/participating users, the host s_j-an activity e_iThe relationship weight of (A) is:

a3, calculating the weight of host-user relationship, and letting SN (u)_i) For user u_iThe number of hosts added, EN(s)_j) As a host s_jThe number of events ever held, SEN (u)_i) For user u_iHosts s participating once_jNumber of events, host s_j-user u_iThe relationship weight of (1) is:

a4, calculating the weight of the user-activity relationship, let E (u)_j) For user u_jSet of activities, s, participated in_jIs an activity e_iThe host, ES (u)_j) For user u_jTake part in_jThe set of held events, then user u_j-an activity e_iThe relationship weight of (1) is:

a5, calculating user-user relationship weight, wherein E (u) is an activity set participated by a user u, E (v) is an activity set participated by a user v, S (u) is a host set participated by the user u, and S (v) is a host set participated by the user v, the relationship weight of the user u-user v is as follows:

a6, and iterative calculation of influence of the host, the user and the activity in the ESU diagram are respectively shown as formulas (1-1), (1-2) and (1-3):

S_k＝c(W_SEE_k-1+W_SUU_k-1)+(1-c)se_j (1-1)

U_k＝c(W_SUS_k-1+W_UUU_k-1)+(1-c)ue_j (1-2)

E_k＝c(W_SES_k-1+W_UEU_k-1)+(1-c)ee_j (1-3)

se in formulae (1-1), (1-2), (1-3)_j、ue_jAnd ee_jRespectively, the jth dimension is a unit vector of 1, and 1-c are return nodes se_j、ue_jAnd ee_jProbability of (W)_SE、W_SU、W_UUAnd W_UEHost-event, host-user, user-user, and user-event relationship weights, respectively.

2. An active social network user according to claim 1A multi-factor decision-making method for activity recommendation, wherein said step (B) of computing an activity e_iAnd user u_jAttended campaign content relevance c (e)_i,u_j) The method comprises the following specific steps:

b1, user u_jThe set of activities that have been attended is denoted E (u)_j) Computing similarity of activities using type labels of activities, CF (e)_i) Representing an activity e_iType tag set of CF (E (u)_j) Represents user u)_jA set of type tags that have participated in the activity;

b2, Activity e_iAnd E (u)_j) The similarity calculation method comprises

3. The method of claim 1, wherein the activity e is calculated in step (C)_iAnd user u_jAttended event location relevance l (e)_i,u_j) The method comprises the following specific steps:

C1、

representing an activity e_iThe location of (a) is determined,

representing user u_jThe location of (a) is determined,

to represent

And

the distance of (d);

c2, Activity e_iLocation and user u_jThe method for calculating the relevance of the place comprises the following steps:

4. the method of claim 1, wherein the activity e is calculated in step (D)_iAnd user u_jTemporal relevance of attended events t (e)_i,u_j) The method comprises the following specific steps:

d1 associating the user has participated in with the activity e_iThe same type of activity set is denoted as E (u)_j) Each activity has a type tag, Activity e_iThe holding time is recorded as

Same type of activity e that the user has participated in_k'∈E(u_j) The holding time is recorded as

Then activity e_iTime of and same type of activity e_kThe time offset of' is:

d2, Activity e_iWith user u_jThe time deviation calculating method comprises the following steps:

min represents taking the minimum value.

Images