Research Of Collaborative Filtering Recommendation Algorithm Based On Hypergraph Convolution

With the rapid development of information technology,we have to deal with the problem of inefficient use of information while enjoying the convenience brought by the Internet.With the huge amount of information,an effective recommendation algorithm can not only provide useful information for users,but also bring great economic benefits to the industry.There are two problems in the existing recommendation algorithms.First,the dataset of the recommender system is based on the user's click behavior.However,due to casual click or exaggerated headlines and pictures,false data will be generated in the process of data collection,which limits the effect of the recommendation model.Second,the user behavior data collected is limited,the collaborative filtering recommendation algorithm is difficult to capture the higher-order correlations between users and items from a small amount of training data.The main contributions of this thesis are as follows:To solve the problem of false data in the recommender system dataset,a concept of the attractiveness and an attractive data collection method are proposed.Attractiveness is defined as the ratio distribution of the gaze point landing in the item area when users click items.The false data can be filtered on the basis of attractiveness between users and items.In the proposed attractive data collection method,the face and eye images are extracted frame by frame from the user video and fed to the i Tracker model to obtain the estimated gaze point.Results demonstrate that the error estimation of the gaze point is stable at 1.8cm,which meets the accuracy requirement of attractive data in actual scenes.To solve the problem of how to fully learn the higher-order correlations between users and items from a small amount of training data,an attention mechanism-based hypergraph convolutional neural network collaborative filtering model(AHGCF)is proposed.The AHGCF model constructs the hypergraph groups to model higher-order correlations between users and items.The vertex embedding is updated by the hypergraph convolutional network.Also,the Lambda Rank algorithm is introduced to supervise the recommendation list.Compared to the best-performing benchmark model Light GCN,the recall and accuracy of AHGCF model is improved by 9.21% and 4.16%.AHGCF model can achieve better performance with a small amount of training data.Several sets of comparison experiments are designed to explore the impact of each module and attractive data on the performance of the AHGCF model.The experimental results show that the model can achieve good performance with a single convolutional layer.The ranking supervision module resulted in a 5.29% improvement in the ndcg metric of the AHGCF model.By introducing the attractiveness data into the AHGCF model in the form of weights,the model is improved 5.18% in recall and 2.18% in ndcg.The effectiveness of the recommendation model is significantly improved.