Research And Implementation Of Graph Neural Network-Based Personlized Recommendation Algorithm

The era of big data led to exponential growth of information on the internet,which has changed the way that people receive information,use it and disseminate it.Due to their limited ability of processing information,explosive information has brought about the problem of "information overload".In recent years,with the development of neural network technology,deep learning based recommendation system,which has ability to process large-scale data processing,extract and combine characteristic automatically,balance memory and generalization,has grown into the growth engine of the Internet.However,there are a variety of data types and complex data representations in recommendation system,which make it difficult to transformer problems into simple processing of sequence data or European spatial data in such fields as natural language processing and computer vision.The emergence of graph neural networks has brought about a deep learning method for processing data in non-Euthic space,which can graph data more efficiently.At the same time,the data structure can effectively organize the key data used in the recommendation system.Therefore,in the framework of deep learning based multi-stage recommendation system,graph neural network is applied as a mining method to explore its feasibility and realization,which is a broad exploration space and full of utilization value.Based on graph neural network,this paper investigates the recommended system and proposes an algorithm to solve existing problems of the multi-stage recommendation system:First,in the matching stage,design a deep learning method,which is based on graph data and sequence data,to complete the matching task.In this paper,the recommended matching model-GCNTREC,combined with Transformer,is proposed.Based on the converse neural network construction diagram information extractor,the model fuses the recommended user-item two-part diagram node neighbor information and generates a dense vector expression of the node,which can enhance the recommendation effect by incorporating the diagram information expression in the deep recommendation model with Transformer as the sequence feature extractor.Second,during the recommended sorting phase,the CTR estimation task is accomplished by effectively utilizing graph type aids and collaborative filtering.This paper proposes the algorithm of multi-task collaborative filtering,GAM-CF,which is enhanced by the graph self-encoder.By designing the multi-task learning method,the model introduces the neural network based reconstruction task of graph of the user and item auxiliary information to the traditional collaborative filtering recommendation task.Then,the model using multi-task learning to optimize multiple loss functions at the same time to make complementary task sharing features,thus improving the idea of original task performance and improving CTR estimation performance.Third,this paper not only builds an intuitive visualization platform for algorithm experiments,but also complete the design and implementation of lightweight simulation recommendation system.The system builds Web services with React and Flask as the architecture.Besides,the system applies Redis,MySql and other databases to organize data in combination with business needs,and integrates the deep recommendation system services,effectively brings GCNTREC and GAM-CF and other algorithm based services online.In this paper,extensive experiments of proposed algorithm is conducted on several open data sets,and the experimental results show that both GGCTREC and GAM-CF can achieve performance improvement compared to the benchmark algorithm on open data sets.Lightweight recommended simulation meets the need for early landing testing of recommended algorithms in an intuitive way.