Research On Recommendation System Based On User Behavior And Time Series

The prevalence of artificial intelligence and big data technology in the Internet era is driving the rapid development of e-commerce websites.Users can obtain relevant information based on their needs on the website,but with it comes the problem of information overload.The emergence of personalized recommendation systems perfectly solves the problem of how to accurately and timely push items that meet user preferences to target users in the vast amount of data,thereby reducing user search time and improving search efficiency.Because most existing recommendation algorithms treat users’ historical behaviors equally,ignoring the differences in user historical behaviors and the inability to selectively process data,resulting in slow recommendation efficiency,further leading to the homogenization of recommendations.In addition,many recommendation systems ignore the issue of user preference changes over time,as well as existing recommendation algorithms that increase in sequence length and undergo iterations,resulting in certain data sparsity and forgetfulness,leading to a decline in recommendation efficiency and quality.Therefore,in response to the above issues,this article focuses on the differences in user historical behavior and implicit information of user historical behavior from the perspective of user behavior and time series,and indepth studies the impact of changes in spatiotemporal information on recommendation systems.Finally,relevant algorithms and models are proposed.The specific research work is as follows:(1)In order to solve the problem of recommendation homogeneity caused by the failure of most sequential recommendation algorithms to mine implicit comment text information,this thesis proposes a new comment text extraction model.The model learns the deep embedded representation of comment text at the user level and item level,and adds aspect level features and implicit feature representations on this basis,enabling the model to more accurately represent user and item features.At the same time,it uses user social relationships to establish a trust model and calculate user similarity.By inputting the final embedded representation of text and social relationships into Trans FM for scoring prediction,a recommendation algorithm based on comment text and social relationships is proposed.Experiments were conducted on Yelp and Epions datasets,and the results showed that the proposed algorithm had a decrease in both RMSE and MAE evaluation indicators compared to the comparison algorithm.(2)Currently,most recommendation models are still in the early stage of integrating spatiotemporal context information,and their processing effects on long sequence data are not ideal.Therefore,this thesis adds spatiotemporal information to the gated cyclic neural network and proposes an improved gated cyclic unit algorithm to capture long-term preferences,while using attention mechanisms to capture shortterm preferences.By fusing the obtained user long-term preference embedding,user short-term preference embedding,and user portrait feature embedding into a multilayer perceptron to predict the next recommendation location,a sequential recommendation model integrating user preferences and spatiotemporal information is proposed.Experimental results on the Foursquare and Brightkite datasets show that the proposed model outperforms the comparison algorithm in HR@K,NDCG@K and MAP@K Three evaluation indicators have been improved.There are 22 figures,14 tables,and 84 references in this thesis.