A method for predicting user online behavior by integrating social group information

Abstract

The present invention discloses a method for predicting user online behavior by integrating social group information. The method first constructs a social network based on the social relationship between users, and constructs an interactive network based on the relationship between users and behavior objects; then reconstructs the social network, and mines the groups contained in the social network; then initializes the users, behavior objects, and mined social groups with randomized low-dimensional vectors; and constructs a user-group information fusion module to obtain the representation vectors of the users and social groups and the final representation vector of the users; then obtains the representation vector of the behavior object; then performs a scoring value for the behavior; and finally uses a Bayesian personalized ranking method to learn and optimize the parameters set in the above model to achieve user behavior prediction. The present invention reduces the impact of noise in user behavior on predicting user behavior from the perspective of group modeling, and improves the accuracy and robustness of user behavior prediction.

Description

Technical Field

[0001] The present invention belongs to the field of user modeling and relates to a user online behavior prediction method integrating social group information. Background Art

[0002] User online behavior prediction is to use data analysis, machine learning and artificial intelligence technology to predict user behavior and decision-making trends in the Internet environment. Its background stems from the acceleration of digital transformation and the increasing complexity of Internet user behavior. In today's digital economy, users interact frequently with platforms, and data is diverse and dynamic. How to provide personalized services and optimize platform strategies through in-depth understanding and accurate prediction of user behavior has become a core issue of concern to various industries. There are many application scenarios for user behavior prediction. First, user online behavior prediction is of great significance to public opinion. By analyzing and predicting user behavior on platforms such as social media, it can timely identify public opinion hotspots, communication trends and emotional trends, and provide support for public opinion guidance and risk resolution. At the same time, it can evaluate the influence of public opinion, optimize communication strategies, identify and control information pollution and rumor propagation, and purify the public opinion environment. Secondly, user online behavior prediction can help companies gain insight into user needs and optimize user experience. For example, by predicting users' purchase intentions, click tendencies or viewing preferences, companies can formulate more targeted recommendations and marketing strategies, thereby improving conversion rates and user satisfaction. In addition, in the fields of security and risk control, user behavior prediction is also widely used in anomaly detection and risk warning, enhancing the security and stability of the system.

[0003] In the data generated by user behavior scenarios, observable user behaviors are generated by the interweaving and joint action of multiple factors. For example, user interest preferences, exposure of behavior objects (items, information), the impact of the environment on users, behavioral norms of user groups, etc. Existing research usually models entity representation based on partially explicit entity relationships, lacks analysis of relationship reliability, and cannot be comprehensively considered from the perspective of user cognition and behavioral motivation. Starting from the multidimensional relationship between entities, the present invention explores various group environmental factors contained in multidimensional social relationships, thereby comprehensively analyzing and modeling users and behavior objects. Without loss of generality, from the perspective of online shopping platforms, when a user buys a product, the user's interest preferences will first be affected by the online group environment in which he is located (such as group influence, social group behavioral norms, etc.). Secondly, there is a lot of noise in the observable user interaction records (such as purchases, clicks, listening, etc.), that is, not all historical records are a true reflection of user interest preferences. For example, a user may listen to a song recommended by others, and after listening to it, the user is likely to dislike the song. However, the system has already left a record of the user listening to the song. If the recommendation system adopts this information later, it is likely to produce an incorrect recommendation.

[0004] How to mine the social group environment to which the user belongs, model the influence of the social group on the user, and then accurately predict the user's behavior is an urgent problem to be solved in the field of user modeling. Based on the social group environment to which the user belongs, the present invention integrates the mutual influence relationship between the user and the social group environment, and proposes a user online behavior prediction method integrating social group information. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings and defects of the prior art, and provide a user online behavior prediction method that integrates social group information. By mining the social groups to which users belong, simulating the iterative influence relationship between social groups and users, and constructing a user-group influence propagation model, the vector representation of users and behavior objects that integrate social group information is realized, and the user's online behavior is predicted.

[0006] The above purpose is achieved through the following technical solutions:

[0007] A method for predicting user online behavior by integrating social group information, the method comprising the following steps:

[0008] Step 1: Use the social relationships between users to build a social network, and use the relationship between users and behavior objects to build an interactive network;

[0009] Step 2: Social network reconstruction: Use the TF-IDF algorithm to calculate the weight value of neighbor users; use the Jaccard coefficient of the user's social neighbor set to calculate the similarity between two users. When calculating the number of users in the neighbor user set, use the weight value of the neighbor user for weighted counting; set the threshold , when the similarity value between users is less than the threshold, the edge relationship in the social network is removed; when the similarity value between users is greater than the threshold, the edge relationship in the social network is retained, and finally a reconstructed social network is obtained;

[0010] Step 3, social group mining: in the social network obtained in step 2, a set of social groups contained in the social network is obtained, and the set of social groups is considered to be the social groups to which the users in the social network belong;

[0011] Step 4: Initialize users, behavior objects, and social groups: initialize vectors of users, behavior objects in social networks and interactive networks, and the social group set obtained in step 3;

[0012] Step 5: Iterative fusion of user-group information based on attention mechanism: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user, and then combine the user's own information to obtain the user representation vector that integrates the social group information. From the perspective of social groups, the representation vectors of users in each social group and the information of the social group itself are integrated to obtain the representation vector of the social group. Through iterative calculation, the mutual influence relationship between users and social groups is simulated to obtain the final representation of users in social networks ;

[0013] Step 6: Fusion of high-level information of user preferences: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain a representation vector of the fused user behavior object information , and the behavior representation vector ; Then multiply the representation vectors of the behavior objects by two products to obtain high-order information between the behavior objects; set the weights , To integrate the above two aspects of user information, we can get the characterization vector of behavior in the interactive network. ;

[0014] Step 7. Behavior score calculation: use weights , The final representation of users in the social network and the user representation in the interactive network are integrated to obtain the user representation that integrates social group information and user preference information. ; Calculate user representation Representation vector of behavior in interaction network The product of is used to obtain the user's score for the target behavior.

[0015] Step 8: Model training: Use the Bayesian personalized ranking method to design the objective function, and use the stochastic gradient descent method to train the model to obtain the model parameters.

[0016] Furthermore, the process of step 2 specifically includes the following steps:

[0017] Step 2.1: Compare users to documents, and compare neighbor users connected by social relationships to words in the document. Use the TF-IDF algorithm to calculate the weights of neighbor users. The TF-IDF value is recorded as ;

[0018] Step 2.2: Calculate the Jaccard coefficient of the social neighbor set between users according to the following formula to measure the similarity between users ;

[0019] ,

[0020] in, Is the user The set of neighbors in a social network, For users The set of neighbors in a social network, Indicates user TF-IDF value of

[0021] Step 2.3: Set the threshold , when the similarity value between users is less than the threshold When the similarity value between users is greater than the threshold, the edge relationship in the social network is removed. When , the edge relationships existing in the social network are retained, and the edge relationships in the social network are retained / removed according to the following formula, and finally the reconstructed social network is obtained;

[0022] ,

[0023] in, Indicates user and users The edge relationship between them, 0 represents the user and users There is no edge relationship between them, 1 represents the user and users There is an edge relationship between them. To set the threshold.

[0024] Furthermore, in step 3, in the social network obtained in step 2, a set of social groups contained in the social network is obtained. Specifically, given an online social network ,in is the set of users in a social network, It is a set of social groups mined using the k-clique permeation algorithm in social networks. use The indicator matrix 𝐇 represents, represents the number of users in the social network, represents the number of social groups in a social network, , where the elements in 𝐇 The definition is as follows:

[0025] ,

[0026] in, , indicating the social groups, if the user Belong to social groups ,but ,otherwise , it means that the user Not part of a social group .

[0027] Furthermore, in step 3, in the social network obtained in step 2, a set of social groups included in the social network is obtained by clustering user attributes through the "group" function of the social platform or by using a clustering algorithm.

[0028] Furthermore, in step 4, vectors of users, behavior objects in the social network and interactive network, and the social groups obtained in step 3 are initialized, and the user, behavior object, and social group representation vectors are initialized in turn as , , .

[0029] Furthermore, the step 5 specifically includes the following steps:

[0030] Step 5.1: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user and calculate the representation vector of the user in the social environment , the calculation formula is as follows:

[0031] ,

[0032] in, represents the sigmoid activation function, For the social groups, , that is, if the user Belong to social groups ,but ,otherwise , Representing social groups The initialization representation vector of is the attribute dimension transformation parameter matrix to be trained, For social groups For users The influence weight of is calculated using a feedforward neural network with a weight parameter a. The calculation formula is as follows:

[0033] ,

[0034] in, is the activation function, is the attribute dimension transformation parameter matrix to be trained, denotes the natural exponential function; the superscript is the matrix transpose symbol;

[0035] Get the user's representation vector in the social environment After that, the user initialization vector representation is fused Get the final representation vector of the user in the social environment = ,in, Element-wise multiplication of vectors;

[0036] Step 5.2: The representation vector of the social group is aggregated from the final representation vector of the user in the social environment. The specific calculation formula is as follows:

[0037] ,

[0038] in, A social group representation vector representing aggregated user information, , that is, if the user Belong to social groups ,but ,otherwise , Represents users in a social environment The final representation vector, is the attribute dimension transformation parameter matrix to be trained, For users For groups The contribution weight of

[0039] After obtaining the social group representation vector of aggregated user information, the representation vector is initialized by integrating the social groups Get the final social group representation vector = ,in, Element-wise multiplication of vectors;

[0040] Step 5.3: Through iterative calculation, simulate the mutual influence relationship between users and social groups, and update the user's representation vector through multiple iterative calculations between users and groups. ,in is the number of iterations; finally, through the splicing operation Get the final representation of users in the social network .

[0041] Furthermore, the step 6 specifically includes the following steps:

[0042] Step 6.1: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain the user representation vector of the fused behavior object information :

[0043] ,

[0044] in, is the learnable dimension transformation parameter, is the weight parameter, A collection of user historical behaviors The Behavior representation vector;

[0045] Step 6.2: When aggregating historical behavior information at a node, perform element-wise multiplication of the vectors Add high-level behavior information in the form of and set the weight , To integrate the above two aspects of user information, finally, we can get the representation of users in the interactive network. , the calculation formula is as follows:

[0046] ,

[0047] in, It is a hyperparameter that adjusts the weight ratio between user historical behavior information and high-level information;

[0048] Step 6.3: From the perspective of behavior in the interactive network, first, use the random sampling method with replacement to randomly sample the set of users who perform the behavior to obtain the user group with the behavior. , use the graph attention network to fuse the target user attributes and obtain the representation vector of the behavior in the interactive network :

[0049] ,

[0050] in, is the learnable dimension transformation parameter, is the weight parameter, The user collection of the behavior object Middle A user representation vector.

[0051] Furthermore, the step 7 specifically includes the following steps:

[0052] Step 7.1: Using weights , Fusion of the final representation of users in social networks and the representation of users in interactive networks , and obtain the user representation that integrates social group information and user preference information :

[0053] ,

[0054] Step 7.2: Calculate user representation The representation vector of the behavior of the target candidate in the interaction network The vector inner product of , obtains the user's preference value for the candidate information , the specific calculation method is as follows:

[0055] .

[0056] Furthermore, the specific method of step 8 is as follows:

[0057] Detailed division of users based on social relationships: users in social networks are divided into socially similar users and interactively similar users through different social relationships and user similarity relationships in interactive networks; socially similar users are users whose number of common friends exceeds a threshold; interactively similar users are users whose number of the same category of behaviors exceeds a threshold; the behavior object sets corresponding to socially similar users and interactively similar users are and , and the two sets have overlapping parts, that is , assuming that users are more likely to have overlapping behaviors, the scores of users and the behavior objects should also be higher. The target users rank the behaviors as follows:

[0058] ,

[0059] ,

[0060] in, Represents the set of behavior objects corresponding to the user in the dataset, ; Represents the set of behavior objects corresponding to similarly interacting users. ; Represents the set of behavior objects corresponding to socially similar users, That is, for the target user For the behavior objects that already exist in the historical records Rating , should be greater than in the set Rating of the behavior object ,Right now ;and Should be greater than in the set and Rating of the behavior object in and ,Right now and ; According to the above sorting method, the final objective function is expressed as follows:

[0061] ,

[0062] Among them, Θ is the parameter in the model, is the regularization term, 𝜆 is the regularization parameter, is the 2-norm, Indicates user Behavior Object The prediction score of Indicates user Behavior Object The prediction score of Indicates user Behavior Object The prediction score of Indicates user Behavior Object The predicted score is trained based on the above loss function using the Adam optimization algorithm to obtain the parameters that make the model perform best.

[0063] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0064] In the prediction of user online behavior, the present invention pays attention to the influence of user interest preferences and behavior objects on the potential social group environment in social networks, and designs a method for predicting user online behavior that integrates social group information. This method combines user historical behavior information and constructs a user representation method that integrates social group information. At the same time, the model incorporates the multi-dimensional correlation between users and historical behavior objects, thereby deeply exploring the user's possible future behavior and improving the accuracy of prediction.

[0065] When modeling users and behavior objects, the present invention designs an iterative influence mechanism between users and their social groups, and incorporates temporal semantic information into the user social network space and user interaction space, thereby achieving vectorized representation of users and objects, and further enhancing the model's ability to capture and express social group information. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of the method of the present invention.

[0067] Figure 2 is an example diagram of a social group of the present invention, Figure 2 (a) shows the community structure group, (b) shows the user-centered network group, and (c) shows the interest group.

[0068] Figure 3 It is a user online behavior prediction model diagram integrating social group information of the present invention. DETAILED DESCRIPTION

[0069] The technical solution of the present invention is described in detail below.

[0070] The following is only one embodiment of the present invention. The present invention may have other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention. These corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

[0071] A method for predicting user online behavior by integrating social group information of the present invention comprises the following steps:

[0072] Step 1: Use the social relationships between users to build a social network, and use the relationship between users and behavior objects (e.g., purchased items, forwarded information, etc.) to build an interaction network;

[0073] Step 2: Social network reconstruction: Use the TF-IDF algorithm to calculate the weight value of neighbor users. TF-IDF (termfrequency–inverse document frequency) is a commonly used weighting technology for information retrieval and data mining. TF is term frequency, and IDF is inverse document frequency. Use the Jaccard coefficient of the user's social neighbor set to calculate the similarity between two users. When calculating the number of users in the neighbor user set, use the weight value of the neighbor user for weighted counting. Set the threshold , when the similarity value between users is less than the threshold, the edge relationship in the social network is removed; when the similarity value between users is greater than the threshold, the edge relationship in the social network is retained, and finally a reconstructed social network is obtained;

[0074] Step 3, social group mining: in the social network obtained in step 2, a set of social groups contained in the social network is obtained, and the set of social groups is considered to be the social groups to which the users in the social network belong;

[0075] Step 4: Initialize users, behavior objects, and social groups: initialize vectors of users, behavior objects in social networks and interactive networks, and the social group set obtained in step 3;

[0076] Step 5: Iterative fusion of user-group information based on attention mechanism: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user, and then combine the user's own information to obtain the user representation vector that integrates the social group information. From the perspective of social groups, the representation vectors of users in each social group and the information of the social group itself are integrated to obtain the representation vector of the social group. Through iterative calculation, the mutual influence relationship between users and social groups is simulated to obtain the final representation of users in social networks ;

[0077] Step 6: Fusion of high-level information of user preferences: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain a representation vector of the fused user behavior object information , and the behavior representation vector ; Then multiply the representation vectors of the behavior objects by two products to obtain high-order information between the behavior objects; set the weights , To integrate the above two aspects of user information, we can get the characterization vector of behavior in the interactive network. ;

[0078] Step 7. Behavior score calculation: use weights , The final representation of users in the social network and the user representation in the interactive network are integrated to obtain the user representation that integrates social group information and user preference information. ; Calculate user representation Representation vector of behavior in interaction network The product of is used to obtain the user's score for the target behavior.

[0079] Step 8: Model training: Use the Bayesian personalized ranking method to design the objective function, and use the stochastic gradient descent method to train the model to obtain the model parameters.

[0080] The process of step 2 in this embodiment specifically includes the following steps:

[0081] Step 2.1: Compare users to documents, and compare neighbor users connected by social relationships to words in the document. Use the TF-IDF algorithm to calculate the weights of neighbor users. The TF-IDF value is recorded as ;

[0082] Step 2.2: Calculate the Jaccard coefficient of the social neighbor set between users according to the following formula to measure the similarity between users ;

[0083] ,

[0084] in, Is the user The set of neighbors in a social network, For users The set of neighbors in a social network, Indicates user TF-IDF value of

[0085] Step 2.3: Set the threshold , when the similarity value between users is less than the threshold When the similarity value between users is greater than the threshold, the edge relationship in the social network is removed. When , the edge relationships existing in the social network are retained, and the edge relationships in the social network are retained / removed according to the following formula, and finally the reconstructed social network is obtained;

[0086] ,

[0087] in, Indicates user and users The edge relationship between them, 0 represents the user and users There is no edge relationship between them, 1 represents the user and users There is an edge relationship between them. To set the threshold; by reconstructing the original social network, users are associated through similar relationships, which enhances the group homogeneity of users in the social network and reduces the noise in the network.

[0088] In step 3 of this embodiment, in the social network obtained in step 2, a set of social groups contained in the social network is obtained. Specifically: given an online social network ,in is the set of users in a social network, It is a set of social groups mined using the k-clique permeation algorithm in social networks. use The indicator matrix 𝐇 represents, represents the number of users in the social network, represents the number of social groups in a social network, , where the elements in 𝐇 The definition is as follows:

[0089] ,

[0090] in, , indicating the social groups, if the user Belong to social groups ,but ,otherwise , it means that the user Not part of a social group .

[0091] In step 3 of this embodiment, in the social network obtained in step 2, a set of social groups contained in the social network is obtained, which is obtained by clustering user attributes through the "group" function of the social platform or by using a clustering algorithm.

[0092] In step 4 of this embodiment, vectors of users, behavior objects in the social network and interactive network, and the social groups obtained in step 3 are initialized, and the user, behavior object, and social group representation vectors are initialized in sequence as , , .

[0093] In this embodiment, step 5 specifically includes the following steps:

[0094] Step 5.1: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user and calculate the representation vector of the user in the social environment , the calculation formula is as follows:

[0095] ,

[0096] in, represents the sigmoid activation function, For the social groups, , that is, if the user Belong to social groups ,but ,otherwise , Representing social groups The initialization representation vector of is the attribute dimension transformation parameter matrix to be trained, For social groups For users The influence weight of is calculated using a feedforward neural network with a weight parameter a. The calculation formula is as follows:

[0097] ,

[0098] in, is the activation function, is the attribute dimension transformation parameter matrix to be trained, denotes the natural exponential function; the superscript is the matrix transpose symbol;

[0099] Get the user's representation vector in the social environment After that, the user initialization vector representation is fused Get the final representation vector of the user in the social environment = ,in, Element-wise multiplication of vectors;

[0100] Step 5.2: The representation vector of the social group is aggregated from the final representation vector of the user in the social environment. The specific calculation formula is as follows:

[0101] ,

[0102] in, A social group representation vector representing aggregated user information, , that is, if the user Belong to social groups ,but ,otherwise , Represents users in a social environment The final representation vector, is the attribute dimension transformation parameter matrix to be trained, For users For groups The contribution weight of

[0103] After obtaining the social group representation vector of aggregated user information, the representation vector is initialized by integrating the social groups Get the final social group representation vector = ,in, It is the element-wise multiplication of vectors, and its purpose is to filter out the effective information of group representation in social network space;

[0104] Step 5.3: Through iterative calculation, simulate the mutual influence relationship between users and social groups, and update the user's representation vector through multiple iterative calculations between users and groups. ,in is the number of iterations; finally, through the splicing operation Get the final representation of users in the social network .

[0105] Step 6 in this embodiment specifically includes the following steps:

[0106] Step 6.1: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain the user representation vector of the fused behavior object information :

[0107] ,

[0108] in, is the learnable dimension transformation parameter, is the weight parameter, A collection of user historical behaviors The Behavior representation vector;

[0109] Step 6.2: When aggregating historical behavior information at a node, perform element-wise multiplication of the vectors Add high-level behavior information in the form of and set the weight , To integrate the above two aspects of user information, finally, we can get the representation of users in the interactive network. , the calculation formula is as follows:

[0110] ,

[0111] in, It is a hyperparameter that adjusts the weight ratio between user historical behavior information and high-level information;

[0112] Step 6.3: From the perspective of behavior in the interactive network, first, use the random sampling method with replacement to randomly sample the set of users who perform the behavior to obtain the user group with the behavior. , use the graph attention network to fuse the target user attributes and obtain the representation vector of the behavior in the interactive network :

[0113] ,

[0114] in, is the learnable dimension transformation parameter, is the weight parameter, The user collection of the behavior object Middle A user representation vector.

[0115] Step 7 in this embodiment specifically includes the following steps:

[0116] Step 7.1: Using weights , Fusion of the final representation of users in social networks and the representation of users in interactive networks , and obtain the user representation that integrates social group information and user preference information :

[0117] .

[0118] Step 7.2: Calculate user representation The representation vector of the behavior of the target candidate in the interaction network The vector inner product of , obtains the user's preference value for the candidate information , the specific calculation method is as follows:

[0119] .

[0120] The specific method of step 8 in this embodiment is as follows:

[0121] Detailed division of users based on social relationships: users in social networks are divided into socially similar users and interactively similar users through different social relationships and user similarity relationships in interactive networks; socially similar users are users whose number of common friends exceeds a threshold; interactively similar users are users whose number of the same category of behaviors exceeds a threshold; the behavior object sets corresponding to socially similar users and interactively similar users are and , and the two sets have overlapping parts, that is , assuming that users are more likely to have overlapping behaviors, the scores of users and the behavior objects should also be higher. The target users rank the behaviors as follows:

[0122] ,

[0123] ,

[0124] in, Represents the set of behavior objects corresponding to the user in the dataset, ; Represents the set of behavior objects corresponding to similarly interacting users. ; Represents the set of behavior objects corresponding to socially similar users, That is, for the target user For the behavior objects that already exist in the historical records Rating , should be greater than in the set Rating of the behavior object ,Right now ;and Should be greater than in the set and Rating of the behavior object in and ,Right now and ; According to the above sorting method, the final objective function is expressed as follows:

[0125] ,

[0126] Among them, Θ is the parameter in the model, is the regularization term, 𝜆 is the regularization parameter, is the 2-norm, Indicates user Behavior Object The prediction score of Indicates user Behavior Object The prediction score of Indicates user Behavior Object The prediction score of Indicates user Behavior Object The predicted score is trained based on the above loss function using the Adam optimization algorithm to obtain the parameters that make the model perform best.

[0127] Simulation experiment:

[0128] The experimental environment of this invention is a PC equipped with Ubuntu 22.04 64-bit operating system, GPU 3090, and memory capacity of 32G. Finally, experiments were conducted on two public datasets (Yelp and Douban). The statistical information of the two datasets is as follows:

[0129] Table 1 Statistics

[0130] Yelp Dataset Douban Dataset Number of users 43,151 17,692 Number of behavior objects 11,952 10,937 Number of interactions 1,348,263 3,251,844 Number of social connections 988,237 1,673,661

[0131] Table 2 Comparison results between the proposed method and other methods on two sets

[0132]

[0133] The above experiments show that the method of the present invention effectively removes redundant noise information in the network by reconstructing the social network, so that the model focuses on the social relationships that really affect user decisions; with the help of the user-group information fusion module constructed by the attention mechanism, the model can automatically identify and weight important social group information, further improving the accuracy of user characterization; at the same time, through the deep mining of the user preference high-order information fusion module and the behavior object information fusion module, the potential associations in the interactive network are fully captured. The experimental results show that this method performs well on different data sets, especially on the Douban data set with strong social attributes, showing higher accuracy and robustness, providing strong support for accurate user behavior prediction.

Claims

1. A method for predicting user online behavior by integrating social group information, characterized in that: The method comprises the following steps: Step 1: Use the social relationships between users to build a social network, and use the relationship between users and behavior objects to build an interactive network; Step 2: Social network reconstruction: Use the TF-IDF algorithm to calculate the weight value of neighbor users; use the Jaccard coefficient of the user's social neighbor set to calculate the similarity between two users. When calculating the number of users in the neighbor user set, use the weight value of the neighbor user for weighted counting; Set a threshold τ. When the similarity value between users is less than the threshold, remove the edge relationship in the social network. When the similarity value between users is greater than the threshold, retain the edge relationship in the social network. Finally, get the reconstructed social network. Step 3, social group mining: in the social network obtained in step 2, a set of social groups contained in the social network is obtained, and the set of social groups is considered to be the social groups to which the users in the social network belong; Step 4: Initialize users, behavior objects, and social groups: initialize vectors of users, behavior objects in social networks and interactive networks, and the social group set obtained in step 3; Step 5: Iterative fusion of user-group information based on attention mechanism: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user, and then combine the user's own information to obtain the user representation vector that integrates the social group information. From the perspective of social groups, the representation vectors of users in each social group and the information of the social group itself are integrated to obtain the representation vector g of the social group k Through iterative calculation, the mutual influence relationship between users and social groups is simulated to obtain the final representation of users in social networks Step 6: Fusion of high-level information of user preferences: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain a representation vector of the fused user behavior object information And the behavior representation vector v j ; Then multiply the representation vectors of the behavior objects by two products to obtain high-order information between the behavior objects; Set the weight γ, 1-γ to integrate the above two aspects of user information and obtain the characterization vector of behavior in the interactive network The specific steps include: Step 6.1: From the user's perspective in the interactive network, use the graph attention network to fuse the behavior object information associated with the user to obtain the user representation vector of the fused behavior object information Among them, σ(·) represents the sigmoid activation function, W4 is the learnable dimension transformation parameter, μ ik is the weight parameter, v j A collection of user historical behaviors The jth behavior representation vector in ; Step 6.2: When aggregating historical behavior information at the node, add high-order behavior information in the form of vector element-wise product ⊙, and set weights γ, 1-γ to fuse the above two aspects of user information. Finally, the representation of the user in the interactive network is obtained. The calculation formula is as follows: Among them, γ is a hyperparameter that adjusts the weight ratio between user historical behavior information and high-order information; Step 6.3: From the perspective of behavior in the interactive network, first, use the random sampling method with replacement to randomly sample the set of users who perform the behavior to obtain the user group with the behavior. Use the graph attention network to fuse the target user attributes and obtain the representation vector of the behavior in the interactive network. Among them, W5 is a learnable dimension transformation parameter, ξ jk is the weight parameter, u k The user collection of the behavior object The kth user representation vector in ; Step 7, behavior score calculation: Use weight α, 1-α to fuse the final representation of the user in the social network and the user representation in the interactive network to obtain the user representation u that integrates social group information and user preference information i ; Calculate user representation u i Representation vector of behavior in interaction network The product of is used to obtain the user's score for the target behavior. Step 8: Model training: Use the Bayesian personalized ranking method to design the objective function, and use the stochastic gradient descent method to train the model to obtain the model parameters.

2. The method for predicting user online behavior by integrating social group information according to claim 1, characterized in that: The process of step 2 specifically includes the following steps: Step 2.1: Compare users to documents, and compare neighbor users connected by social relationships to words in the document. Use the TF-IDF algorithm to calculate the weights of neighbor users. i The TF-IDF value of i ); Step 2.2: Calculate the Jaccard coefficient of the social neighbor set between users according to the following formula to measure the similarity between users sim(u i ,u j ); in, Is user u i The set of neighbors in a social network, For user u j The set of neighbors in a social network, T(u k ) represents user u k TF-IDF value of Step 2.3: Set a threshold τ. When the similarity value between users is less than the threshold τ, remove the edge relationship in the social network. When the similarity value between users is greater than the threshold τ, retain the edge relationship in the social network. Retain / remove the edge relationship in the social network according to the following formula, and finally obtain the reconstructed social network. Among them, e ij Represents user u i and user u j The edge relationship between them, 0 represents user u i and user u j There is no edge relationship between them, 1 represents user u i and user u j There is an edge relationship between them, and τ is the set threshold.

3. A user online behavior prediction method integrating social group information according to claim 1 or 2, characterized in that: Step 3: In the social network obtained in step 2, a set of social groups contained in the social network is obtained, which is obtained by clustering user attributes using the "group" function of the social platform or by using a clustering algorithm.

4. A user online behavior prediction method integrating social group information according to claim 1 or 2, characterized in that: Step 3: In the social network obtained in step 2, obtain the set of social groups contained in the social network. Specifically, given an online social network g = (V, g), where V is the set of users in the social network, G is the set of social groups mined in the social network using the k-clique percolation algorithm, and the social network It is represented by the indicator matrix H of |V|×|G|, where |V| represents the number of users in the social network, |G| represents the number of social groups in the social network, |V|>>|G|, and the elements in H are The definition is as follows: Among them, g j ∈G, represents the jth social group in the social network. If user u i Belong to social group j ,but 1, otherwise This means that user u i Not part of a social group j .

5. The method for predicting user online behavior by integrating social group information according to claim 4, characterized in that: Step 4 initializes the vectors of users, behavior objects in the social network and interactive network, and the social groups obtained in step 3, and initializes the user, behavior object, and social group representation vectors to u in turn. i , v j , g k .

6. The method for predicting user online behavior by integrating social group information according to claim 5, characterized in that: The step 5 specifically comprises the following steps: Step 5.1: From the user's perspective in the social network, use the attention mechanism to integrate the influence of the social group on the user and calculate the representation vector of the user in the social environment The calculation formula is as follows: Among them, σ(·) represents the sigmoid activation function, g k is the kth social group, That is, if user u i Belong to social group k ,but otherwise g k Indicates social group k The initialization representation vector, W1 is the attribute dimension transformation parameter matrix to be trained, α ik For social groups k For user u i The influence weight of is calculated using a feedforward neural network with a weight parameter a. The calculation formula is as follows: Where LeakyReLU is the activation function, W2 is the attribute dimension transformation parameter matrix to be trained, exp(·) represents the natural exponential function, and the superscript T is the matrix transposition symbol. Get the user's representation vector in the social environment After that, the user initialization vector representation u is fused i Get the final representation vector of the user in the social environment Among them, ⊙ is the vector element-wise multiplication; Step 5.2: The representation vector of the social group is aggregated from the final representation vector of the user in the social environment. The specific calculation formula is as follows: Among them, g′ k A social group representation vector representing aggregated user information, That is, if user u i Belong to social group k ,but otherwise Represents user u in a social environment i The final representation vector, W3 is the attribute dimension transformation parameter matrix to be trained, β ki For user u i For group g k The contribution weight of After obtaining the social group representation vector of aggregated user information, the representation vector g is initialized by integrating the social groups k Get the final social group representation vector Among them, ⊙ is the vector element-wise multiplication; Step 5.3: Through iterative calculation, simulate the mutual influence relationship between users and social groups, and update the user's representation vector through multiple iterative calculations between users and groups. Where l is the number of iterations; finally, the final representation of the user in the social network is obtained through the concatenation operation || 7. The method for predicting user online behavior by integrating social group information according to claim 1, characterized in that: The step 7 specifically includes the following steps: Step 7.1: Use weights δ, 1-δ to fuse the final representation of users in the social network and the representation of users in interactive networks Get user representation that integrates social group information and user preference information Step 7.2: Calculate user representation The representation vector of the behavior of the target candidate in the interaction network The vector inner product of the user's preference value for the candidate information is obtained. ij , the specific calculation method is as follows:

8. The method for predicting user online behavior by integrating social group information according to claim 7, characterized in that: The specific method of step 8 is as follows: Detailed division of users based on social relationships: Based on different social relationships and user similarity relationships in the interactive network, users in the social network are divided into socially similar users and interactively similar users; socially similar users refer to users whose number of common friends exceeds a threshold; interactively similar users refer to users whose number of the same category of behaviors exceeds a threshold; the behavior object sets corresponding to socially similar users and interactively similar users are V SS and V S , and the two sets have overlapping parts, that is Assuming that users are more likely to have overlapping behaviors, the scores between the user and the behavior object should also be higher. The target user ranks the behaviors as follows: f:r ip ≥r in ；r in ≥r ik ；r ij ≥r im v p ∈V p ;v n ∈V SS ∩V S ;v k ∈V S ;v m ∈V SS Among them, V p Represents the set of behavior objects corresponding to the user in the dataset, v p ∈V p ; V S represents the set of behavior objects corresponding to similar interactive users, v k ∈V S ; V SS represents the set of behavior objects corresponding to socially similar users, v m ∈V SS ; That is, for the target user u i For the behavior object v that already exists in the history p Rating ip , should be greater than in the set V SS ∩V S The rating of the behavior object r in , that is, r ip ≥r in ; and r ij Should be greater than in the set V S and V SS The rating r of the behavior object in ik and r im , that is, r in ≥r ik and r in ≥r im ; According to the above sorting method, the final objective function is expressed as follows: Among them, Θ is the parameter in the model, is the regularization term, λ is the regularization parameter, ‖Θ‖2 is the 2-norm, Represents user u i For behavior object v o The prediction score of Represents user u i For behavior object v n The prediction score of Represents user u i For behavior object v k The prediction score of Represents user u i For behavior object v m The predicted score is trained based on the above loss function using the Adam optimization algorithm to obtain the parameters that make the model perform best.

Images