Research And Implementation Of Sequence Recommendation System Based On Deep Learning

With the rapid development of the Internet industry,society has entered an era of information explosion,and recommendation systems based on anonymous session click sequences have become a current research hotspot.In the context of increasingly strict regulations on the acquisition of personal information on the Internet by national policies,if relatively few clicks information can be used in anonymous sessions to make effective personalized real-time recommendations,it will help companies achieve the goal of reducing costs and increasing efficiency.Therefore,this paper conducts research on anonymous session-based click sequence recommendation.The main research contents and achievements are as follows:(1)An anonymous session click sequence recommendation model based on a multi-layer parallel graph neural network is proposed.The model uses a multi-layer and scalable parallel graph neural network with attention mechanism to carry out parallel modeling and learning on the features of the click sequences.Finally it can fully perform the cross-fusion of explicit and implicit features,simulating the decision-making process of user in real scenarios,so as to improve the recommendation effect.In addition,the model adopts the mechanism of sharing parameters between parallel graph neural networks,which reduces the training parameters to improve the training speed.The model is tested and compared with multiple benchmark models on two classic e-commerce public datasets.The experimental results show that the model has improved in both MRR@20 and P@20 indicators.The model training speed is improved after using parameter sharing mechanism.(2)The high-performance deep learning model deployment and inference framework called FastDeploy is proposed and implemented.The framework utilizes cross-platform compilation related technologies,integrates multiple backend inference engines and introduces highperformance computing libraries,so as to solve the problems that the existing frameworks cannot be deployed across platforms,the backend inference engine support is not comprehensive,the deployment model format is not uniform,and the multi-platform inference performance drops.In addition,FastDeploy Server,which is an industrial-grade highperformance service-oriented deployment framework,is implemented by using FastDeploy combined with the Triton Inference Server framework.(3)Design and implement a high performance anonymous session click sequence recommendation system.The system uses big-data related technologies Spark and Flink to extract features,uses in-memory database Redis and server memory cache to store features hierarchically,uses local sensitive hash algorithm to improve recall speed,and uses FastDeploy Server for model loading and serving deployment of the sorting layer algorithm called SR-PGNN.Finally,a high-performance recommendation system that can make real-time recommendations based on session click sequences is realized.