Research And Implementation On Recommendation System Based On Feature Fusion

For the past few years,more and more researches began to apply the idea of feature fusion to recommendation system.However,the existing recommendation algorithms based on feature fusion still have some defects.First,the current mainstream recommendation algorithm lack global feature fusion model design and only fuse the feature information of part of the model.Secondly,the existing algorithms based on self-supervised learning and feature fusion only focus on the pre-training task,and fail to further utilize and mine the implicit feature information in the data.The thesis studied the above problems and proposed new solutions.The following are the primary contents of the thesis.:(1)A self-supervised sequential recommendation model S~3Rec-UT is proposed.The model divides user behavior sequences according to the sequential interval characteristics,distinguishing long-term and short-term user behavior sequences.On this basis,the model screens out two data enhancement methods and designs corresponding pre-training tasks.By jointly modeling the long-term and short-term user behavior sequences with two pre-training tasks,a vector representation that fuses user timing features and context features is obtained.Experiments on three Amazon datasets show that compared with the optimal baseline model,the HR@10 indicators increased by 1.4%,1.3%and 0.7%,respectively,and the NDCG@10 indicators increased by 2.4%,1.2%and 0.3%.(2)A global feature fusion recommendation framework GF3Rec is proposed.From the perspective of global features,the model designs a hierarchical fusion structure.Through the three hierarchical structures of feature embedding fusion layer,context feature extraction layer,and gating feature fusion layer,the model obtains a more comprehensive and accurate representation of user characteristics in a progressive fusion manner.Experiments in two real data set scenarios have proved the effectiveness of the model.Compared with the optimal baseline model,the HR@10 indicators have increased by 1.2%and 3.3%,respectively,and the NDCG@10 indicators have increased by 1.5%and 4.8%.(3)Combining the above two recommendation algorithms,this thesis designs and develops a personalized product recommendation system based on the MVC architecture.Design and plan three structures of view layer,model layer and data layer according to the requirements of requirement analysis,and realize the development of each functional module in accordance with the principle of interface-oriented development in the development process.After testing,all functional modules of the system can operate normally,and can successfully complete the task of personalized recommendation.