A recommendation method based on social network user behavior information

Abstract

The application relates to a recommendation method based on social network user behavior information, which comprises the following steps: obtaining user behavior data and obtaining user activity holding location preference and holding time preference; constructing a preference similar network by using the obtained user preference, and then fusing a user online social relationship network to obtain a plurality of non-overlapping community structures reflecting the social connection and preference similarity between users; calculating the value of each hidden point representing the user's decision on a project by using a hybrid pairwise Markov random field model; calculating all node states of each project, and then selecting a recommendation list by sorting. The application aims to recommend the optimal activity holding location and holding time for an activity organizer, so as to maximize the number of participants in the activity and meet the decision planning needs of the activity organizer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of social network recommendation system and data mining technology, and particularly relates to a recommendation method based on social network user behavior information. BACKGROUND

[0002] An activity social network that combines online and offline interaction: In the activity social network, an activity organizer publishes social activities such as academic conferences and outdoor activities online, and publishes details of the activities, including information such as the theme, category, scale, organizer, offline holding time and offline holding location of the activities. Social network users select activities of interest by browsing and searching online, and participate in the activities offline at the holding time of the activities. In this way, the online and offline combination breaks the limitation of traditional social networks that only have online social functions, and users can communicate with users having common interests and make friends when participating in activities offline.

[0003] In the activity social network, the number of participants in an activity is crucial to the success of the activity, so for an activity organizer, how to maximize the number of participants in the scale of the activity is an important goal of the organizer. The theme, holding location and holding time of an activity are important factors that affect whether a user participates in an activity. When an activity organizer organizes an activity, the biggest challenge is how to select the holding location and holding time of the activity to maximize the number of participants.

[0004] Currently, researchers focus on different needs of users in the activity social network and propose various methods, such as activity recommendation methods, activity participant prediction methods, or plan a series of activities to be held by an activity organizer, consider the conflicts between activities, and how to allocate these activities to different times to maximize the number of participants, given a set of activities, a set of activity holding times and a set of users. Although these studies provide decision support for users and activity social network platforms, helping users quickly obtain activities of interest and improving their experience on the platform, few studies start from the perspective of activity organizers to recommend optimal activity holding locations and holding times to help them make decisions and planning to maximize the number of participants in the activity. To solve this problem, the present application proposes a joint recommendation method for activity holding locations and times to provide decision support for activity organizers. SUMMARY

[0005] Technical problems solved

[0006] The present application aims to provide a joint recommendation method for activity holding location and time based on the behavior information of social network users from the perspective of activity organizers, and provide decision support for activity organizers.

[0007] Technical scheme

[0008] To achieve the above object, the present application is implemented by the following technical scheme:

[0009] The present application provides a recommendation method based on the behavior information of social network users, which is used to recommend n activity holding location-time period pairs for an activity to be held, wherein the definition of location-time period pair is a project , and the recommendation method comprises:

[0010] S10, obtaining data information in a social network platform, including user behavior data, a set containing multiple activity holding locations, a set containing multiple activity time periods, and an activity theme;

[0011] S20, constructing a user-location matrix and a user-time matrix according to the obtained data information, and modeling the user's location preference and the user's time preference by fusing matrix decomposition and network embedding method, respectively, to obtain the user's preference for the activity holding location and the preference for the holding time;

[0012] S30, constructing a preference similarity network by using the obtained user preference, and obtaining multiple non-overlapping community structures reflecting the social contact and preference similarity between users by fusing the user online social relationship network;

[0013] S40, calculating the value of each representative user point for the project decision by using a hybrid pairwise Markov random field model, wherein when the implicit point = 1, it indicates that the user will participate in the activity held in the location-time period pair;

[0014] S50, calculating all node states of each project , and counting the number of values of 1, i.e. the number of users participating in the activity, and then sorting the number of participating users of each project from high to low, and selecting the first n projects to constitute a recommendation list, which is used for high-quality and efficient selection by an activity decision maker.

[0015] Further, in step S20, the steps of constructing the user's location preference and the user's time preference modeling are as follows:

[0016] S21: constructing a user-location matrix and a user-time matrix according to the obtained data information, respectively;

[0017] S22: performing matrix decomposition on the user-region matrix and the user-time matrix respectively;

[0018] S23: constructing a user-region bipartite graph and a user-time bipartite graph based on user historical behavior data, and generating user-region embedding representation and user-time embedding representation through random walk and Node2Vec method;

[0019] S24: fusing the latent factors obtained by matrix decomposition and the embedding representation to construct a preference model, and using the preference model to calculate the final preference of the user in the region and the time.

[0020] Further, in step S24, an optimization objective function is set, which combines the loss of matrix decomposition and the loss of network embedding, wherein the optimization objective function about user region preference is specifically:

[0021]

[0022] wherein,

[0023] In the above formula, is the objective function, , , is a regularization parameter, and are true value and predicted value respectively, is the embedding representation of the user in the region, is the latent preference of the user in the region, the latent feature of the region , the embedding representation of the region .

[0024] Further, in step S24, when calculating the final preference, the latent factor is fused with the embedding representation to obtain the final preference of the user in the region , the latent factor is fused with the embedding representation to obtain the final preference of the user in the region .

[0025] Therefore, the comprehensive preference of the user in the region and the time is

[0026] Further, the latent factor is fused with the embedding representation , obtaining the latent features of the region , fusing the latent factors , with the embedding representation , obtaining the latent features of the time , ; Finally, the latent features of the item

[0027] are obtained. Further, in step S30, the community structure is obtained, including the following steps:

[0028] S31: based on the obtained user preference information, the preference similarity between users is calculated, and the calculation formula is as follows:

[0029]

[0030]

[0031] A similarity threshold is set , if the preference similarity between users , an edge is established between users , otherwise, there is no edge between users ; by calculating the preference similarity between any users, a social network based on user preferences is obtained.

[0032] S32: the user online social relationship network is converted into an undirected graph, and then the new user online social relationship network is fused with the preference similarity network to obtain a single-layer network.

[0033] S33: the single-layer network is divided into multiple non-overlapping community structures by using the Louvain algorithm based on modularity.

[0034] Further, in the mixed pairwise Markov random field model, two types of nodes, observation nodes and hidden nodes, are included, edges are constructed between the hidden nodes, and the edges represent the dependency relationship between the nodes. The dependency relationship includes two types, one is the relationship between the same user, and the other is the relationship between the same item. For the same item, if two users are located in the same community structure, there is an edge between them.

[0035] Further, in step S50, when calculating the value of the hidden point of each item i,

[0036] the hidden node ​​​​​​​​​and the potential function definition is carried out on the edge, wherein, representing a user making a decision on a project ;

[0037] representing a user making a decision on a project

[0038] ; representing a user making a decision on a project

[0039] ;

[0040]

[0041] wherein,

[0042] when represents that the user will participate in the activity held in the region and the time period , represents that the user will not participate in the activity held in the region and the time period .

[0043] beneficial effects

[0044] The technical scheme provided by the application has the following beneficial effects compared with the known prior art:

[0045] The scheme is innovative based on the historical behavior information of the user, acquires the location preference and time preference of the user through the preference model, and reflects multiple non-overlapping community structures of the social contact and preference similarity between the users, can generate an activity holding region-time period pair recommendation table for an activity planner, and recommends the optimal activity holding place and holding time for an activity organizer for a social network to be held, so as to maximize the number of participants of the activity, meet the decision planning needs of the activity organizer, reduce the decision difficulty of the activity planner, and effectively improve the quality of the activity holding. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only show some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0047] Figure 1 The step flow chart for the recommendation method of the present application;

[0048] Figure 2 The step flow chart for the preference modeling of the present application;

[0049] Figure 3 The step flow chart for obtaining the community structure of the present application;

[0050] Figure 4 The user-area matrix decomposition schematic diagram of the present application;

[0051] Figure 5 The application schematic diagram of the hybrid pairwise Markov random field model of the present embodiment; DETAILED DESCRIPTION

[0052] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only show some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0053] The present application will be further described below in combination with the embodiments.

[0054] Embodiment:

[0055] The present application provides a recommendation method based on social network user behavior information, the object of the present recommendation method is an activity planner, and the recommendation problem is: for an activity social network, given a social activity to be held , a candidate area set and a candidate time period set and the historical behavior information of the user, the recommendation problem based on social network user behavior information is defined, and the target is to maximize the number of participants of the activity, and the activity organizer is recommended for an activity to be held to recommend an activity holding area and a holding time period , that is, , particularly, for convenience, a region-time period pair is called an item , for convenience, a region-time period pair is defined as an item ,

[0056] Referring to Figure 1 , the method for recommending the region r and the time period t of an event includes:

[0057] S10, obtaining data information including user behavior data, a set R including a plurality of event holding locations, a set T including a plurality of event time periods, and an event theme X in a social network platform;

[0058] S20, constructing a user-region matrix and a user-time matrix according to the obtained data information, and modeling the region preference of a user and the time preference of the user by fusing a matrix decomposition method and a network embedding method, to obtain the location preference of the user and the time preference of the user, respectively;

[0059] S30, constructing a preference similarity network by using the obtained user preferences, and obtaining a plurality of non-overlapping community structures reflecting the social connection and the preference similarity between users by fusing a user online social relationship network;

[0060] S40, calculating the value of each representative user for the implicit point of an item by using a hybrid pairwise Markov random field model, where the implicit point = 1 indicates that the user will participate in the event held in the region-time period pair;

[0061] S50, calculating all node states of each item , and counting the number of values equal to 1, i.e., the number of users participating in the event, and then sorting the number of participating users of each item from high to low, and selecting the first n items to form a recommendation list, which is used for the event decision maker to make a high-quality and efficient selection.

[0062] The scheme innovatively obtains the location preference and the time preference of a user by a preference model based on historical behavior information of the user, simultaneously obtains a plurality of non-overlapping community structures reflecting the social connection and the preference similarity between users, and establishes a hybrid pairwise Markov random field model, so as to generate an event holding region-time period pair recommendation list for an event planner, to recommend the optimal event holding location and the optimal event holding time for an event organizer of a social network to be held, to maximize the number of participants of the event, to reduce the decision difficulty of the event planner, and to effectively improve the quality of the event holding.

[0063] In this embodiment, all the geographic locations are divided into a set of candidate regions based on the locations of all the activities in the social network platform using the k-means clustering algorithm , i.e. , wherein, represents a certain region. At the same time, the time of all activities is divided into a set of candidate time periods , in particular, 21 time periods are defined, 7 time periods with day interval (each time period represents a day of the week) and 3 time periods with time interval (morning, afternoon and evening), i.e. T=Monday morning, Monday afternoon… Sunday afternoon, Sunday evening. Based on the above, the location of each activity is quantified as a region , and the time of each activity is quantified as a time period . Finally, each activity in the activity social network corresponds to an activity theme , an activity region , and an activity time period .

[0064] The preference of a user is a key factor affecting whether the user participates in a social activity, therefore, based on the behavior information of the user in the activity social network, modeling the preference of the user is the key to solving the social network recommendation problem, in step S20, referring to Figure 2 , the steps of constructing the user's regional preference and the user's time preference modeling are as follows:

[0065] S21: According to the obtained data information, construct a user-region matrix and a user-time matrix , each element in the user-region matrix represents whether the user has participated in an activity held in the region in the past,

[0066] each element in the user-time matrix represents whether the user has participated in an activity held at time in the past.

[0067] S22: Perform matrix decomposition on the user-region matrix and the user-time matrix respectively;

[0068] S23: Construct a user-region bipartite graph and a user-time bipartite graph based on the user historical behavior data, and generate user-region embedding representation and user-time embedding representation through random walk and Node2Vec method;

[0069] S24: fuse the latent factors and the embedding representation obtained by matrix decomposition to construct a preference model, and use the preference model to calculate the final preference of the user in the region and time.

[0070] In steps S22-S24, the embodiment is illustrated by taking the modeling of the region preference of the user as an example, and the algorithm principle of the time period preference modeling of the user is consistent.

[0071] Specifically, based on the constructed user-region matrix, a method of fusing matrix decomposition and network embedding is designed to more accurately obtain the region preference of the user. The specific steps are as follows:

[0072] 1. Based on the user-region matrix , the matrix decomposition method is used to decompose it into a user region preference matrix and a region attribute matrix , refer to Figure 4 , wherein represents the latent region preference of the user , and represents the latent attribute of the region . The optimization objective function of matrix decomposition is:

[0073]

[0074] 2. Network embedding representation learning based on the user-region matrix again. First, based on the user-region matrix constructed based on the historical behavior information of the user, a user-region bipartite graph is constructed, in which the behavior data of the user is composed of a binary tuple , if the user participates in the activity held in the region , then an edge is established between the user and the region .

[0075] Secondly, based on the constructed user-region bipartite graph, a random walk is used to walk any node in the bipartite graph, and the walk will be repeated until a predetermined step length is reached. In this way, a random walk can obtain a path that captures the network structure and deep semantics. The walk probability of the random walk is:

[0076]

[0077] wherein represents the neighbor set of the node in the bipartite graph.

[0078] Then, based on the random walk path, node sequences with different starting nodes are obtained. Furthermore, nodes are identified by their co-occurrence with nodes within a fixed-length window. Construct neighborhood set The Node2Vec method, using a Skip-Gram approach, is used to generate embedded representations of nodes. The optimization objective function is:

[0079]

[0080] in, Represents a node Negative samples.

[0081] Furthermore, based on the aforementioned objective function, stochastic gradient descent is used to optimize this objective, learning the nodes. Embedded representation Finally, based on the above steps, the user's... Embedding on a region is represented as ,area The embedding is represented as

[0082] 3. To obtain more accurate user region preferences, a preference calculation method integrating matrix factorization and network embedding is designed. In step S24, an optimization objective function is set, which combines the loss from matrix factorization and the loss from network embedding. Specifically, the user region preference optimization objective function is as follows:

[0083]

[0084] in,

[0085] In the above formula, Let be the objective function. , , It is a regularization parameter. and Separate the actual value and the predicted value. For the user's embedded representation on the region, For users' potential preferences in a region, and the potential characteristics of that region. Region embedding representation Ultimately, the latent factors are integrated. With Embedded Representation Get users Final preference in the region .

[0086] The integrated preference calculation, which combines regional and temporal factors, is performed in step S24 when calculating the final preference, incorporating latent factors. with embedding representation , obtaining the final preference of user on region , fusing latent factors with embedding representation , obtaining the final preference of user on region ; thus, obtaining the comprehensive preference of user on region and time as ; further fusing latent factors with embedding representation , obtaining the latent features of region , fusing latent factors with embedding representation , obtaining the latent features of time ; finally, obtaining the latent features of item .

[0087] Further, in step S30, the community structure is obtained, referring to Figure 3 , comprising the following steps:

[0088] S31: based on the obtained user preference, calculating the preference similarity between user and user , the calculation formula is as follows:

[0089]

[0090] setting a similarity threshold , if the preference similarity between user and user is , then an edge is established between user and user , otherwise, there is no edge between user and user ; by calculating the preference similarity between any users, a social network based on user preference, i.e. a preference similarity network, is obtained;

[0091] S32: converting the user online social relationship network into an undirected graph, then fusing the new user online social relationship network with the preference similarity network to obtain a single-layer network;

[0092] S33: using Louvain algorithm based on modularity to divide the single-layer network into multiple non-overlapping community structures.

[0093] ​​​By using the probabilistic inference ability of Markov random field, the hybrid pairwise Markov random field model can more effectively improve the accuracy of recommendation;

[0094] With reference to Figure 5 In the hybrid pairwise Markov random field model, two types of nodes are included, i.e., observation nodes and hidden nodes. Circular nodes are observation nodes, and hexagonal nodes are hidden nodes. For example, as an observation node, the node represents the preference of a user , represents the attribute feature of an item , and as a hidden node represents the decision of a user on an item , and takes a value of 0 or 1. Each hidden node is connected to two observation nodes and , and the value of each hidden node is calculated by a user preference modeling module. The state of each hidden node is determined by the observation nodes connected to it (i.e., the user preference and the item feature ) and the hidden nodes connected to it (i.e., the node , the node and the node ).

[0095] In the model, edges are constructed between hidden nodes, and the edges represent the dependency relationship between nodes. The dependency relationship includes two types. One type is the relationship between users, such as the edge between the node and the node in the figure. The other type is the relationship between items. For the same item, if two users are located in the same community structure, an edge exists between them, such as the edge between the node and the node in the figure, and the edge between the node and the node .

[0096] In step S50, when calculating the value of the hidden node of each item i,

[0097] first, the potential function is defined for the hidden node and its edges, where represents the decision of a user on an item . Specifically, for the node , represents the node potential function, and represents the potential function of its edge. Wherein, is the node After given the user 's preference and the item 's , the prior probability distribution is calculated by Cosine similarity between ,

[0098]

[0099]

[0100] For the edge between node and , its potential function reflects the similarity between item and , which is calculated by,

[0101]

[0102]

[0103] For the edge between node and , its potential function reflects the similarity between user and located in the same community, which is calculated by,

[0104]

[0105]

[0106] The joint probability distribution of the hybrid pairwise Markov random field is represented as Gibbs distribution;

[0107] The joint probability distribution of the hybrid pairwise Markov random field is represented as,

[0108]

[0109] Wherein, is the normalization factor, and are the user set and edge set of the hybrid pairwise Markov random field.

[0110] For each node , the belief propagation algorithm is used to infer its posterior probability ; ​

[0111] The posterior probability of each node in the mixed pairwise Markov random field model is calculated The state of each node is determined by the following calculation,

[0112]

[0113] Wherein,

[0114] When represents that the user will participate in the event held in the area and the time period , represents that the user will not participate in the event held in the area and the time period . Then, in order to achieve the goal of maximizing the number of participants of an event, the state of each node is calculated for each item , and the number of nodes with a state of 1, i.e. the number of users participating in the event is calculated. Finally, the number of participating users of each is sorted from high to low, and the first few items i are selected to form a recommendation list and pushed to the project planner.

[0115] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or can be implemented by means of software plus necessary general hardware platforms. Based on such understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the recommendation method described in each embodiment of the present application.

[0116] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present application.

Claims

1. A recommendation method based on social network user behavior information, wherein the recommendation method is used to recommend n pairs of regions-time periods for an upcoming event, wherein, Define a region-time period pair as a project The recommendation method is characterized by comprising: S10, Obtain data from social networking platforms including: user behavior data, sets containing multiple event locations, sets containing multiple event time periods, and event theme data; S20, construct user-region matrix and user-time matrix based on the acquired data information, and combine matrix decomposition and network embedding methods to model user regional preferences and user time preferences respectively, and obtain user preferences for event location and event time respectively; S30: By using the acquired user preferences to construct a preference similarity network, and then integrating the user's online social relationship network, multiple non-overlapping community structures that reflect the social connections and preference similarities among users are obtained; S40 utilizes a hybrid pairwise Markov random field model to calculate the pairwise relationships between each representative user and the item. The value of the implicit point of the decision; when the implicit point = 1, it means that the user will participate in the activities held in this area-time period. S50, calculate each item The status of all nodes is recorded, and the number of nodes with a value of 1 is counted, representing the number of users participating in the activity. Then, each activity is sorted from highest to lowest. The number of participating users determines the selection of the top n projects. This constitutes the recommendation list; In step S20, the steps for constructing the user's regional preferences and time preferences model are as follows: S21: Construct user-region matrices based on the acquired data. and user-time matrix ; S22: For the user-region matrix respectively and user-time matrix Perform matrix decomposition; S23: Construct user-region bipartite graphs and user-time bipartite graphs based on user historical behavior data, and generate user-region embedding representations and user-time embedding representations through random walk and Node2Vec methods; S24: The latent factors and embedding representations obtained from matrix decomposition are fused to construct a preference model, and the preference model is used to calculate the user's final preferences in terms of region and time; The hybrid pairwise Markov random field model includes two types of nodes: observed nodes and hidden nodes. Edges are constructed between the hidden nodes, and the edges represent the dependencies between the nodes. There are two types of dependencies: one is the relationship between the same user, and the other is the relationship between the same project. For the same project, if two users are located in the same community structure, then there is an edge between them.

2. The recommendation method based on social network user behavior information according to claim 1, characterized in that, In step S50, when calculating the value of the hidden point for each item i, First, target the hidden nodes. The potential function is defined for its edges, where, On behalf of users For the project Decision-making; The joint probability distribution of the mixed pairwise Markov random fields is represented as a Gibbs distribution; For each node The belief propagation algorithm is used to infer its posterior probability. ; The calculation yields each node in the mixed pairwise Markov random field model. After calculating the posterior probability, its state is determined by the following calculation: ; in, , when Indicates user Will participate in the region and time period The events held Indicates user Will not participate in the region and time period The events held.

Images