Research On Conversational Recommender System Based On Sparse Graph Attention And Location Perception

With the rapid development of Internet technology and the rapid expansion of the global digital economy,the amount of data is growing exponentially.In this era,personalized recommendation,as an effective information filtering method,has gradually attracted the attention of researchers.Traditional recommendation algorithms rely on users’ historical behavior records to perceive their interest preferences,and then provide recommendations.As a new research direction that combines dialogue systems and recommendation systems,conversational recommendation algorithms can obtain users’ interest preferences in real time through natural language interaction.However,dialogue texts lack sufficient contextual information,and the conversational recommendation model may not be able to accurately give appropriate responses based on the current situation.In recent years,knowledge graphs have been proven to be an effective way to introduce prior knowledge for high-level semantic relationship mining as auxiliary information fusion into conversational recommendation systems.In view of the problems existing in current knowledge graphbased conversational recommendation systems in various scenarios,this paper designs a more efficient and accurate recommendation algorithm to optimize user experience.The research content of this paper is summarized as follows:(1)In most cases,graph convolutional neural networks being unable to filter taskirrelevant nodes in knowledge graphs and the long dependency problem of recurrent neural networks,this paper proposes a conversational recommendation algorithm based on multihop attention and sparse graph attention mechanisms.Firstly,an external vector is introduced as a memory unit in this paper,and the multi-hop attention mechanism is used to encode the dialogue,with constant updates being made to the external vector.Furthermore,it models the semantic information of the knowledge graph,and uses a binary mask to mark the edges in the graph to distinguish noise nodes in the knowledge graph.In addition,KL annealing optimization is used to optimize the variable autoencoder,solving the problem of KL divergence disappearance.The proposed method not only effectively alleviates the long dependency problem but also improves the model’s ability to capture spatial information and reduces computational costs.(2)This paper proposes a conversational recommendation algorithm based on additive attention and positional encodings to address the issue that existing knowledge graph-based conversational recommendation algorithms usually can’t effectively utilize the topology of the graph,and thus cannot fully leverage the higher-order semantic information of the graph.Firstly,the model uses additive attention mechanism to encode text and transforms pairwise interactions between different words in a sentence into modeling of global context and different words.In addition,to fully utilize the position information of nodes,the model replaces the position constant with Laplacian eigenvectors and extends the application of transformer to graph data,expanding local attention to global attention.The proposed method not only effectively reduces the computational complexity of transformer but also captures the topological information of the graph,achieving the expansion of the receptive field of graph neural networks.(3)A recommendation system application platform based on additive attention and position-aware graph neural networks has been designed.This application platform enables real-time interaction with users through Streamlit’s text control.The movie dialogue interface is deployed on an Nvidia Ge Force RTX 3090 server using the Sanic web service framework.Additionally,the interaction records between the platform and users are continuously written into log files,providing valuable data for further analysis and improvement of system performance.With this platform,users can easily access the movie recommendation system and receive personalized recommendations based on their preferences and feedback.The platform also offers a user-friendly interface and efficient processing capability,providing a seamless and enjoyable experience for users.To some extent,it enhances the efficiency of information retrieval for users.