Research And Implementation Of Recommendation Algorithm Combined With Graph Neural Network

In the era of mobile Internet,the emergence of massive amounts of information not only brings a larger and more comprehensive view to users but also brings the problem of information explosion.How to help users quickly and efficiently screen out their required information from complex information has become an urgent problem for all Internet companies.The emergence of neural network technology and big data has made it possible to extract information from massive data and make use of it.As a result,the recommendation system based on big data came into being and quickly became an important growth engine for Internet companies.The data structures in the recommendation system are diverse and complex.In addition to the regular Euclidean spatial data,there are also non-Euclidean spatial data similar to the graph structure.These structures are difficult to exploit but contain a wealth of information.The emergence of graph neural network makes it possible to efficiently extract graph structure information,and its powerful representation ability can introduce recommender systems into a whole new field.Therefore,the introduction of graph neural networks in recommender systems has a very broad prospect and great value.Based on the research on graph neural networks and recommendation algorithms,this paper proposes two recommendation algorithms combined with graph neural networks and a set of engineering systems.At the algorithm level,this paper proposes two improvements:1.A general graph recommendation algorithm HGCNNCL is proposed for the recall and ranking links in the recommender system.The algorithm uses graph convolutional neural network sampling to aggregate high-order neighbor information of users/items in the user-item bipartite graph to generate node representations.On this basis,HGCNNCL introduces the idea of Hypergraph and Contrastive Learning,it strengthens the representation ability of the model by constructing new positive and negative samples and multi-task learning.In terms of the generality of the algorithm,in the recall stage,HGCNNCL can obtain node vector representations through pretraining and use vector recall online;in the sorting stage,HGCNNCL can be slightly transformed into an end-to-end model,and vector training can be improved through direct supervision of downstream tasks.for full.2.Aiming at the situation in the current recommendation system where there is a small amount of interaction between users and products and the user has missing features,a cold-start graph recommendation algorithm GCMCSR based on feature reconstruction is proposed.Performing feature reconstruction tasks can alleviate the cold start problem.In the actual recommendation stage,the algorithm can be added to the service as a cascade model,providing recommendation opinions and supplementing the reconstructed user characteristics to the subsequent recommendation model.At the engineering level,this paper designs and builds a lightweight recommendation system,which can help algorithm developers simulate online recommendation scenarios.The system adopts React and Django as the architecture uses MySQL database for data storage and interaction,and integrates various recall and sorting models,which can help developers to observe the online effect of the model more intuitively.For HGCNNCL,this paper designs a two-stage experiment of recall and sorting.A total of four real-world datasets,five different evaluation indicators,and eleven comparison methods are set up in the two stages.The experimental results show that the algorithm is suitable for all stages of the recommendation process,and the effect is improved compared with the commonly used algorithms in the industry.For GCMCSR,we conduct comparative experiments with five methods on four real-world datasets.At the same time,a feature reconstruction experiment is also designed for the cold start scenario.The two sets of experimental results demonstrate that GCMCSR can improve the recommendation effect while alleviating the cold start problem.Finally,this paper completes the development and launch of a lightweight recommendation system,and implements functions such as personalized recommendation,user behavior analysis,and algorithm effect analysis in a simulated online environment.