| With the improvement of Internet services,the network data available in recommender system is increasingly rich,and with the development of natural language processing and graph neural network technology,recommender system still has a lot of room to improve.Based on this,the main work of this thesis is as follows:(1)In order to enhance the utilizatin of review text features in the ranking stage of recommendation,this thesis proposes a ranking model FCRF based on the dual attention and the full combination of review features by combining the FM’s feature crossover idea and the attention mechanism.FCRF utilizes Bert model to encode review text to enhance the representation of review text,and then uses cross attention and multi-head self-attention to realize the cross combination and self combination of the user’s review features and the item’s review features.The average MSE of FCRF on four real datasets is 1.43%less than the state of the art review-based model DAML.(2)In order to enhance the utilization of user-item interactions and reviews in recommendation system,this thesis proposes a brand new graph embedding training model RHG.Based on the edge-based graph neural network,RHG model utilizes reviews to represent edges in the graph,then use the edges to aggregate neighborhood node information based on different feature spaces.Finally,on four real datasets,RHG improves by 2.49%compared with the state of the art text-based graph embedding training model CGAT,and the trained graph embedding is used to embed into the FCRF model,which reduce the MSE by 0.65%.(3)In order to improve the accuracy of recall stage,this thesis proposes a recall model NFRM based on the graph embedding.The model constructs target nodes’ feature in the feature neighborhood and the spatial neighborhood in the bipartite graph of users and items,and selects PCA to perform dimension reduction before feature crossover.The PQ algorithm which is improved by the idea of random forest is also used in NFRM to improve retrieval efficiency of recall.Finally,on three real datasets,NFRM has an average increase of 3.87%in the hit rate compared with the state of the art graph-based recall model EGES.(4)In order to make the recall model and predict model designed in this thesis have more engineering value,this thesis designs a music recommendation website based on Django framework,which uses FCRF and NFRM as predict model and recall model respectively. |
