Research On Quaternion-Based Graph Convolution Network For Recommendation

With the continuous development of mobile Internet technology,the amount of information in the network is increasing exponentially,and the problem of information overload is becoming more and more serious.However,human information processing capabilities are limited,and the amount of information received by users is much higher than what they can handle,which makes it difficult for them to find the content they are interested in from the massive amount of information and seriously interferes with the user’s analysis and selection.In order to alleviate the user’s information overload dilemma and analyze and select the information that meets the user’s preferences from a large amount of overloaded information,recommender systems came into being.Among them,collaborative filtering recommendation is one of the important scenarios in recommender systems.Collaborative filtering recommendation aims to describe the user’s interest preferences based on the user’s historical interaction,so as to recommend more suitable items to the user.In real life,the historical interaction data of users is diverse and complex.How to obtain more accurate embedding representations of users and items is a major problem faced by collaborative filtering recommendation.Based on the graph convolution network and user historical interaction data,this thesis models users and items in the quaternion space,a hypercomplex space,to solve the problem of embedding distortion of users and items.At the same time,in view of the problems of lack,noise,and data sparsity in user historical interaction data,this thesis designs from the model level and the auxiliary task level in the quaternion space to obtain more accurate embedding representations of users and items and improve recommendation performance.The main work of this thesis is as follows:(1)Aiming at the problems of lack and noise in user historical interaction data,this thesis proposes a quaternion-based graph convolution network recommendation method.This method is modeled in the quaternion space,and the message propagation mechanism of graph convolution network in the quaternion space is designed,which effectively alleviates the problems caused by the lack and noise in user historical interaction data.Finally,the effectiveness of the method is verified by experimental comparison.(2)Aiming at the problem of sparse user historical interaction data,this thesis proposes a quaternion-based graph contrastive learning recommendation method.This method is modeled in the quaternion space,and the contrastive learning auxiliary task is constructed to extract the self-supervised signal in the user-item interaction data,which effectively alleviates the problem caused by the sparse user historical interaction data.Finally,the effectiveness of the method is verified by experimental comparison.(3)This thesis designs and implements a product recommender system based on the quaternion-based graph convolution network recommendation method.This system utilizes the quaternion-based graph convolution network to analyze the user’s historical interaction data to obtain the user’s interest preferences and provide personalized recommendation services to the user.