Research On Graph Learning Based Recommendation Algorithm Integrating Multimodal Learning And Self-supervised Learning

With the advent of the Internet era,information overload has become a pressing problem due to the explosive growth and increasing complexity of information,so recommendation systems were proposed in the 1990 s and became a specialized discipline for research.However,traditional recommendation algorithms based on graph neural networks are inadequate in the direction of cold start processing,higher-order information aggregation,and label-free learning.So in this thsis,three algorithmic models are proposed around the multimodal data recommendation problem,using multimodal learning,self-supervised learning,and graph data structure enhancement.Firstly,to address the problems of cold start in recommendation systems and the difficulty of adaptively capturing user preferences in traditional graph neural networks,this thesis proposes a multimodal graph attention network recommendation method based on multimodality for video,audio and text data.The method embed multimodal representations into the model to model user preferences,then introduces attention mechanisms into the graph neural network framework for separating user preferences for different modalities,and further introduces a gated mechanism to control and weight the information flow in the multimodal interaction graph to facilitate the understanding of user behavior.Secondly,to address the problems of high cost of manual annotation,loss of higher-order information in graph convolutional neural networks and weak generalization ability of the model due to the high reliance on labels in recommender systems,this thesis proposes a self-supervised capsule graph convolutional network based recommendation method,which uses capsule graph convolutional networks to construct supervised signals,and employs three data enhancement modes to construct contrast views and self-supervised signals,and further employs Info NCE to construct contrast loss for the introduction of self-supervised learning.Thirdly,this thesis presents an attempt to fuse multimodal and self-supervised learning for graph learning recommendation,aiming to address these limitations of traditional recommendation paradigms that are affected by sparse and untouched patterns and structures of supervised data,and to further explore the potential of recommendation systems.For video,audio and text data,we propose a method for fusing multimodal learning and self-supervised learning graph neural network recommendation in this thesis,which designs a multimodal parallel graph-based recommendation model as the main supervised learning task and three different granularity data augmentation methods to build multimodal self-supervised components according to the multichannel paradigm.Finally,this thesis conducts extensive experiments on datasets such as Tiktok,Kwai,Movie Lens,Last-FM and Yelp,and compares them with state-of-the-art baseline models to explore and investigate key modules and important parameters to analyze and validate the effectiveness and superiority of the proposed models.