Research On Recommendation Algorithm Based On Transformer And Graph Neural Network

With the rapid development of Internet technology,the explosive growth of information has led to a serious information overload problem,making it difficult to discern and select useful information.The purpose of recommendation systems is to solve this problem.A recommendation system is an effective information retrieval tool whose core is a recommendation algorithm.Classical recommendation algorithms model the implicit relationships between users and items through techniques such as collaborative filtering and matrix decomposition.With the development of deep learning,deep neural network-based approaches have become the mainstream choice for recommendation algorithms,and a variety of different neural networks have been applied to recommendation algorithms,including simple multi-layer perceptron,convolutional neural networks,recurrent neural networks,Transformer and graph neural networks,etc.However,these methods still have a series of problems: first,the existing models have a tendency of increasing model complexity,which not only increases the training difficulty but also affects the practical application of the algorithms;second,the existing research neglects the effective use of auxiliary information,such as social networks and knowledge graphs.To address the above problems,this paper hopes to combine the techniques of self-supervised learning,knowledge graph,graph neural network and recommendation algorithm in some way based on the existing research,and to achieve the performance improvement while using only Transformer as the backbone network.The specific research of this paper is as follows:(1)In this paper,a two-stage sequential recommendation method is proposed based on self-supervised learning and pre-training + fine-tuning training paradigm.The algorithm first processes the user behavior sequences as pre-training training samples,pre-trains them on Transformer for the mark detection task,and then fine-tunes them on the recommendation task.The method uses only Transformer as the interest extractor,which avoids the training difficulty problem caused by the complex network structure and not only improves the performance but also accelerates the convergence of the model.(2)In this paper,a simple robustness-enhanced collaborative filtering recommendation method is proposed based on the ideas of self-supervised comparative learning and graph neural networks.The algorithm adds Dropout operation to the graph neural network computation to generate contrast view pairs to assist training,and the model is improved in both performance and convergence speed.(3)In this paper,based on the first two studies,the rich knowledge in the knowledge graph is extracted based on graph neural network technology as auxiliary information to enhance the item representation.The algorithm enhances the performance and mitigates the interference of overfitting phenomenon on the performance.(4)In this paper,the proposed method is extensively experimented and analyzed on several publicly available datasets such as Amazon,Movie Lens,and Yelp.And compared with stateof-the-art models such as Light SANs,CORE,Sim GCL,etc.respectively,the experimental results prove that all three algorithms proposed in this paper have better performance.