Research On Social Influence Prediction Of Graph Neural Network Based On Multi-View And Knowledge Distillation

In the Internet age,as the development of information technology,a variety of social network platforms and applications have emerged in people's daily life.Meanwhile,recently the scale of Internet users has also obviously increased.More and more users begin to share and interact on social network platforms such as Weibo,We Chat and Twitter,which brings large amount of social network data.For researchers,it is of great significance to effectively analyze and utilize these social network information.In social networks,users will have a certain influence on each other through sharing and exchanging information,which is named social influence.To a great extent,social influence will change user's views and behavior decisions.For social networking platforms,advertising recommendation,e-commerce and other applications,more accurate prediction of user behavior can not only provide more effective information but also bring better user experience.Therefore,social influence prediction is a research hotspot in data mining field and has great application value and practical significance.Traditional social influence prediction methods and recent deep learning social influence prediction methods do not consider the different attributes of user in social networks,and how to effectively mine the information of user' s different attributes is worth exploring.To solve this problem,this thesis proposes a deep learning social influence prediction method based on multi-view learning and knowledge distillation.The main work includes the following two aspects:(1)For the difference and diversity of user attributes in social networks,this thesis incorporates multi-view learning into graph attention network.Based on which this thesis proposes a deep learning method to implement an efficient social influence prediction model,named Multi-view Influence prediction network(Mv Inf).Mv Inf takes user's multi-view features as inputs to mine the complementarity and consistency among different views to improve the learning and prediction performance,which can predict more accurately whether the user's behavior in the social network is affected.Experiments on different social network datasets show that the proposed Mv Inf model in this thesis outperforms other social influence prediction methods based on single view.(2)Considering the complexity of the deep learning model,the deep learning method based on multi-view learning tends to be more complex and requires more computing cost.To solve this issue,this paper introduces the knowledge distillation strategy into the multi-view graph attention network,and designs a Social Influence prediction method(Dis Mv Inf)based on distillation graph neural network model.Dis Mv Inf utilizes multi-view models with better performance and more complex structure to guide the learning of single-view models,which can implement a lightweight model with maintaining better prediction performance.Experiments on different social network datasets also show the efficiency and effectiveness of Dis Mv Inf.