Research On Recommendation Algorithms Based On Bayesian Personalized Ranking And Information Dissemination

The popularity of the Internet has made the resources on the network more and more abundant,everyone can become a producer of information,you can get almost everything you want to know from the network.But this also brings a variety of problems,"information explosion" is one of them.To alleviate this problem,recommendation systems have been created.Recommender systems can help people find the content they are interested in from a myriad of information,which makes it a very common tool.Recommender systems are not only convenient for ordinary people to get information,but also widely used in business.Internet companies such as Tiktok,Jingdong,Baidu,and Bilibili are actively using recommender systems to increase user viscosity and improve transaction rates to gain benefits.Unlike previous recommender systems that can only obtain information about first-order neighbors,graph neural network-based recommender systems can aggregate information about higher-order nodes,which makes it have quite a bright performance when recommending.The graph neural network-based recommendation system also has its inherent shortcomings: first problem is over-smoothing,in the training proceeds,the embedding vector of each node in the graph will slowly tend to be the same.This will lead to a gradual decrease in the recommended results when a certain number of training sessions is reached.Second,the speed of training is much slower than traditional recommendation algorithms,because the graph neural network has to multiply the node’s embedding vector matrix and the adjacency matrix each time the information is propagated,and n layers have to be multiplied n times in one training,which leads to the high cost of graph neural network training.To address the above shortcomings,this paper makes the following improvements to the recommendation algorithm based on graph neural networks:(1)The randomly initialized embedding vectors are first fed into a lightweight graph neural network for higher-order information propagation and training.Whenever a certain number of training times is reached then the difference of nodes in the graph in the embedding space is increased using Bayesian personalized ranking matrix decomposition(BPRMF).The number of training times can be adjusted for different datasets so that it can be adapted to different datasets.Finally,the vector inner product is used to predict user preferences and the K items with the highest user preferences are selected for recommendation.Experiments are conducted on two datasets,and the experimental results show that this algorithm can effectively alleviate the oversmoothing in training and improve the recommendation effect with Recall@20 and NDCG@20 as evaluation metrics.(2)Unlike explicit message passing,in Chapter 4 we choose to use constrained loss to directly let the embedding vector approximate the result of infinite layer propagation.This can greatly reduce the time consumption caused by information propagation.Also,using this approach allows for more appropriate assignment of edge weights and more flexibility in adjusting the importance of different types of relationships.Tests are conducted using Recall@20 and NDCG@20 as evaluation metrics in two datasets.The experimental results show that the training speed and recommendation performance are both improved compared with previous graph neural network-based recommendation systems.