Deep Learning Based Research On Online Social Influence And Its Application In Group Recommendation

With the rapid development of online social networks,how to mine users’ social influence in massive interactive information and large-scale social networks,and promote downstream application research such as advertising and recommender systems,has attracted tremnedous interest from academia and industry.In th is paper,we focus on two points:(1)online social network influence prediction at the individual level;(2)group recommendation with integrated influence.We comprehensively analyze influence from two granularities of user and group,and argue that the precise characterization of influence at the user level helps to model the influence of users during group decision-making,so as to recommend items(food,movies or tourist spot,etc.)that most members are satisfied with.The influence prediction task focuses on predicting whether users are likely to be influenced by friends on social networks to interact(forward or like,etc.).Most of the current methods focus on the macro level of information cascade prediction,while the micro level does not fully consider the relationship between users /connection strength between nodes.The group recommendation task aims to model how members of a group reach decisions and recommend items that meet the expectations of most members.Existing work focuses more on the influence of users within a group,while ignoring the difference in the influence of users across groups.In view of these shortcomings,we aim to enhance the existing methods to improve the performance of social influence prediction and group recommendation.The main research work and results of the paper are as follows:First,we propose a method to incorporate connection strength into influence prediction.(1)We construct a connection strength matrix according to the following time duration between users or the number of interactions between users.(2)We design a graph neural network model based on modeling connection strength: the network representation learning method Deep Walk is used to extract the node network structure feature representation,and the connection strength matrix is used to explicitly measure the connection strength between nodes.The aggregation mechanism incorporated with connection strength is designed in the graph convolutional neural network,and the attention mechanism incorporated with connection strength is designed in the graph attention neural network,aiming to model the influence between nodes,and finally the fully connected layer is stacked to predict whether the user may be influenced to interact.(3)We conduct comprehensive experiments on two social network datasets,and the results show the superiority of the method proposed in this paper,and the contribution of connection strength features to the improvement of model performance.Second,we propose a method to incorporate influence into group recommendation.(1)We design a contrastive learning method,inspired by the fact that a user can form a positive sample with the group he is in,and a negative sample with the group not in.(2)We design a multi-task learning model fused with social influence: under the main task of group recommendation,we use the attention mechanism to simulate the mutual negotiation between users during decision-making,aming to capture the variable influence of users in the group,and use the contrastive learning task to model the phenomenon that users have different influences in different groups,and alleviate the cold start issue caused by sparse interaction data between groups and items through user recommendation tasks.(3)We conduct experiments on three public datasets.The results show that user recommendation task,contrastive learning task and attention mechanism contribute to the model performance improvement.