Research On Group Environment-sensitive Social Recommendation

The rapid development of online social networks makes social recommender systems gradually become a hot spot in academia and industry.The social recommendation systems introduce social information into the field of traditional recommendation,aim to mine the correlation between user behaviors based on social psychology theories and practical experiences in social network analysis,attempting to alleviate the data sparsity problem in the recommendation scene and improve the final recommendation performance.At present,a lot of research progress has been made in this field at home and abroad,which provides inspiration and reference for our continued research in the field of social recommendation.However,there are still some deficiencies in this field:Firstly,the influence theory is the basic theory in social recommendation research,but existing social influence research mainly focuses on the influence between individuals,and lacks modeling of the complex social environment to which target users belong.Secondly,the existing social recommendation models are still in the exploratory stage for considering the group environment and dynamics that affect users’ interest preferences.Finally,there is a lack of an evaluation framework for the utility of social information in social recommendation field.To overcome the above weaknesses and challenges,this dissertation conducts a systematic study of social influence analysis and social recommendation.Specifically,the research work of this dissertation is divided into three aspects:·This research proposes a multi-relational group influence model to solve the problem of lacking definition and modeling of group influence in social influence analysis.The model fully considers the group environment to which social users belong and comprehensively quantifies the group influence.Firstly,based on the social psychology theories,we define the rules of group size and group scope and then construct static group environments and dynamic group environments.Secondly,we introduce multi-layer graph attention networks to simulate the influence diffusion in different group environments,respectively.Finally,the influence of users in multiple group environments is integrated to model the group influence in online social networks.The experimental results based on multiple social network datasets demonstrate the rationality and effectiveness of the proposed group influence model in this dissertation.·This research proposes a temporal-aware and multifaceted social contexts-based social recommendation model to dynamically modeling of group environment influence in the social recommendation.The model combines multiple group information and considers the dynamics of user preferences.Specifically,based on the structural characteristics of the social network,the model first disentangles the online social environment into the community structure-based social context and the ego-network-based social context.Secondly,in social contexts,the model learns the higher-order information and the dynamic influence between users.Finally,by integrating information from different groups,the model can accurately infer user preferences.The experimental results on multiple public recommendation datasets demonstrate that the proposed social recommendation model has better performance in all evaluation metrics.·To fill the gap of lacking evaluation metrics of social information in the social recommendation,we propose a social information evaluation framework that can comprehensively evaluate social recommendation models and effectively evaluate the usefulness of social information.Specifically,firstly,we propose three research questions based on the existing research on social recommender systems:(a)How useful is social information in recommender systems?(b)How well does the social recommendation model utilize social information?and(c)how does social information play a role in the cold start problem?Secondly,based on the research question,a social information recommendation evaluation framework is proposed,which consists of a novel ablation analysis experiment and two evaluation metrics related to social information.Finally,based on the proposed evaluation framework,we evaluate and analyze the existing social recommender systems.Through experiments,we find that with the increase of user-item interaction information,the usefulness of social information decreases in the existing social recommendation model.Through the above research,we developed methods for social network analysis theory,a social recommendation model,and an evaluation framework for social recommender systems,which provides theoretical support for the development of social recommendation.In practical applications,this dissertation also provides practical experience for social recommendation in optimizing online resource allocation,reducing user time costs,and improving social resource utilization.