Music Recommendation System Based On Improved Hard Negative Sample And Neural Network

In the era of big data,the popularity of the Internet not only facilitates people’s access to information,but also makes people fall into the dilemma of choosing massive information.Academia and industry have reached an agreement to solve this problem,believing that recommendation systems are one of the most effective ways to alleviate data overload.This article develops a music recommendation system based on improved hard negative samples and neural networks,using two different algorithms,namely,improved neural networks and negative samples,to extract the characteristics of users and music.The feature information is processed through neural collaborative filtering(NCF)and graph convolutional neural networks(GCN)models,and ultimately recommends music that may be of interest to users.In the research on the neural network-based music recommendation model,this paper chooses NCF,which combines multi-layer perceptron and generalized matrix factorization,as the basic model,and integrates it with the current popular contrastive learning(CL)technology to improve the recommendation performance.At present,the recommendation model based on contrastive learning mainly focuses on the improvement and expansion of GCN,but there is relatively little practice on the recommendation algorithm based on neural network.Therefore,this paper proposes a simpler CL method,which is different from graph-based data augmentation methods,but adds uniform noise in the feature space to create contrastive views,and uses CL as part of model training to continuously improve model accuracy.In terms of recommendation algorithms based on improved hard negative samples,this paper selects GCN as the basic research model.As one of the advanced methods of collaborative filtering,GCN effectively mines the information of each node through aggregation operations,improving the performance and accuracy of the system.However,the negative sampling of GCN has uncontrollable randomness,which is not sufficient to reflect the value of negative samples.For this reason,the focus of this study has shifted to improving the negative sampling algorithm for GCN,and an algorithm for generating hard negative samples(Hns),called noise fusion(NF＿GCN),has been proposed.NF＿The GCN algorithm mainly includes three stages: noise interference,fusion strategy,and optimization strategy.In the noise interference stage,a positive sample interference term is generated by adding a certain amount of noise to the positive sample feature.In the fusion strategy stage,positive sample interference terms and negative sample features are mixed to obtain the Hns set.In the optimization strategy stage,by calculating the similarity of each sample in the set,the Hns with the highest score is selected as a negative sample to participate in model training.This article designs and implements a simple music recommendation system for transplanting the two improved algorithms.The system uses the Vue.js framework to design front-end pages,while the back-end uses Django for development and design.The database uses My SQL for data management.