Research On User Behavior Prediction Based Of Social Networking Service Platform

With the advent of the era of big data,the rapid flow of information and data has provided a good environment for the development of the Internet.As a medium for data and user interaction,a large number of Social Networking Services(SNSs)platforms that have been born in recent years have increasingly become an indispensable and important component of life.It not only brings convenience to people's lives and improves user satisfaction,but also contains huge scientific and economic research value to increase considerable economic benefits.In terms of user services,SNSs provide online platforms,on the one hand,the recommendation system,a user-oriented information retrieval technology,is used to obtain the information that users need,and to select products that may be of interest to them,so as to alleviate the problems caused by information overload.On the other hand,the platform uses the topological structure of social networks and user characteristics to find potential social relationships for users.From the perspective of users,user behaviors on the aforementioned SNSs platforms mainly include(1)User social link behaviors that establish social link relationships.(2)Consumption behavior of users browsing or purchasing items of interest.Accurately providing users with effective personalized recommendations,predicting their consumer interest preferences,and mining potential social links have become two core tasks on SNSs platforms and a research hotspot with practical significance.However,the existing behavior prediction algorithm models face problems and challenges such as the sparsity problem,joint modeling of different user behaviors and complex relationship characterization,and low sensitivity for discovering and identifying higher-order social relationships and interest preferences.To this end,this dissertation focuses on user behavior prediction tasks,theoretically combining data mining technology,user behavior,sociology,and other cross-disciplinary knowledge;technically combining machine learning and deep learning technology that has achieved success in multiple fields in recent years,such as network embedding learning,deep neural networks,and graph convolutional networks and practical application scenarios,researches on the above-mentioned problems and challenges.Specifically,the major contributions of this dissertation are as follows:First,we propose a joint user behavior prediction method named RWJBG based on network embedding learning.This method combines the social network formed by user history records and the interest network formed by user consumption records through the same user as a bridge to form a joint behavior graph,which alleviates the sparsity problem.Next,RWJBG uses a random walk strategy to sample each node in the joint behavior graph multiple times to obtain node sequences.In this way,the associations and complex relationships that are difficult to describe between different behaviors are collected,and the node sequence information about the high-order relationships between nodes is obtained.Finally,RWJBG uses the language model Skipgram to learn the feature expression of each node,so that the learned node features can be used to calculate the correlation between nodes and applied to the prediction tasks of two user behaviors.Second,we propose a user joint behavior prediction model called NJBP based on feature fusion and link distance.With the great success of deep learning in computer vision and natural language processing,the technology of deep learning has gradually been applied to the prediction of user consumption behavior with the recommendation system as the technical background,and the social link behavior prediction with link prediction as the technical background.In view of this,we propose a joint neural network model NJBP based on deep learning to jointly predict user behavior to solve the aforementioned problems and challenges.In the use of data,NJBP combined different user behavior data to alleviate the problem of data sparsity.Technically,NJBP uses the nonlinear fitting ability of deep neural networks to break the linear model limitation of traditional matrix factorization algorithms.In addition,NJBP uses different feature fusion methods to characterize and establish the complex connection between the two behaviors,so that the learned features can include the characteristics of social influence and homogeneity assumptions.In order to capture higher-order social relationships,NJBP uses the topology information of social networks during model training to distinguish the impact of social neighbors at different distances.Experimental results show that the introduction of these structures and settings plays a key role in improving the accuracy of social link behavior prediction and consumer behavior prediction.Third,we propose a predictive model of user consumption behavior integrating social and interest network.Because most of the research work of the same kind mainly relies on the user's first-order link relationship to model user behavior,the learned representations need to be improved and optimized in describing the overall network topology information.In fact,the influence of social and interest preferences in the real world may present high-order characteristics of transitivity.In order to model this transfer diffusion process,we propose the model DiffNet++.First of all,because users play a core role in social networks and interest networks,DiffNet++uses the same users in the two networks to construct a heterogeneous graph of social networks and interest networks as inputs to alleviate the impact of data sparsity.Next,user embeddings collect higher-order information from the social network and the interest network through iterative aggregation from heterogeneous graphs,perform feature fusion and update,and simulate the relevance and influence of different behavior records.In addition,in order to accurately distinguish the impact of information aggregation on different users in different networks(social networks and interest networks),and to characterize the differences in information aggregation of different nodes under the same network,we designed a neural network based on the attention mechanism to perform precise modeling.Finally,a large number of experimental results on four real data sets verify the effectiveness of the proposed model.