| With the rapid development of the Internet,the amount of information that people are exposed to has exploded,leading to a demand for effective and targeted information acquisition.Recommendation system has also emerged.The collaborative filtering based recommendation algorithm does not make too many assumptions and restrictions on data,and can handle large-scale and highly sparse data.It has the advantages of universality and flexibility,so it is widely used in e-commerce,social networking,news recommendation,film recommendation and other fields.The graph neural network based collaborative filtering algorithm has advanced performance,so it attract the attention of a large number of researchers.Through the investigation and analysis of the research status,the current collaborative filtering models based on graph neural network still have the following three problems:(1)The recommendation performance of graph convolution based models do not improve with the increase of model complexity:Graph convolution based models do not improve the structure of graph convolution networks for collaborative filtering.Therefore,compared with matrix factorization based models,although the models’ complexity is increased,theirs performance do not significantly improve;(2)The graph convolution based models have oversmooth problem:If the model uses multi-layer graph convolution to extract information,it is equivalent to performing a stacked low-pass filtering operation on the input signal,which limits the model’s ability to express high-order information;(3)Hypergraph based models lack implicit information learning and obtain relatively single information:The hyperedges of hypergraph models are all explicitly constructed,and can only learn information in specific directions.And the information extracted by hypergraph convolution is mostly high-order information,lacking the supplement of low-order information.In order to solve the above problems,this thesis carry out the following work:(1)In response to problem 1,we propose a two-stage matrix decomposition network based on graph convolution named TSNet.Firstly,we integrate the graph convolution into three parts:self-information aggregation module,neighborhood-information aggregation module,and relationship-information aggregation module.Through ablation experiments,we prove that there are redundant structures in the graph convolution.Then,we combine the graph convolution structure with matrix factorization and replace the redundant parts with specialized information aggregation structure.As a result,we reduces the complexity of the graph convolution and make better information extracting;(2)For problems 2 and 3,we propose a collaborative filtering network based on adaptive hypergraph named CFAH.First,through experiments we prove that the model based on graph convolution has the problem of oversmoothing.We analyze this problem in the spatial domain and frequency domain,and then we construct adaptive hypergraph to deal with the collaborative filtering to improve the phenomenon of oversmoothing.Compared with the manually specified hyperedges,the iterative hyperedges can better learn the hidden links in the data.Finally,we combine the hypergraph convolution with the traditional graph convolution,enable the model to extract both high-order and low-order information simultaneously;(3)Through experimental verification on multiple datasets,we demonstrate the proposed method has good recommendation performance.Then we demonstrate the effectiveness of the strategy designed in this paper through ablation experiments. |
