Accuracy And Data Sparsity Oriented Recommendation Methods Research

With the rapid development of the global digital economy and the exponential growth of data,the massive amount of information exceeds the threshold for individual users to process it,thus leading to information overload problems.As an efficient means to solve the information overload,recommendation systems have received high attention in all major mainstream platforms and applications.On the one hand,there is a room to improve the accuracy of recommendation systems when multiple features interfere with decision making;on the other hand,existing methods usually require a lot of labor cost and sacrifice accuracy when solving the data sparsity problem.To address the above problems,this thesis proposes a Recurrent Disentangled Knowledge Graph Convolutional Network for accuracy,and a Self-supervised Knowledge Graph Neural Network for data sparsity problems.The innovation and main work of this thesis are as follows:(1)To address the problem that multiple feature factors jointly interfere with decision making in the real world leading to low accuracy of recommendation systems,this thesis proposes a Recurrent Disentangled Knowledge Graph Convolutional Network.Firstly,the algorithm adopts the feature extraction module to learn the representation of the knowledge graph,and then captures the key features by deeply disentangled multiple feature factors through the steps of disentangled featur,confidence calculation,neighborhood aggregation and self-attention mechanism in the recurrent disentangled feature module,calculates the entity scores according to the score prediction module,and finally obtains the recommendation results after model training.In this thesis,we conduct multiple sets of simulation experiments on Movie Lens-20 M,Book-Crossing and Last.FM datasets,and integrate the recall rate,F1-score,AUC evaluation metrics to verify the effectiveness of the model proposed in this thesis.The experimental results demonstrate that the RDKGCN model achieves a recall rate of 34.48% on the Movie Lens-20 M dataset,which is an improvement of about2.22% compared to the suboptimal model.(2)To address the data sparsity problem caused by massive data,this thesis proposes a Self-supervised Knowledge Graph Neural Network.Firstly,the algorithm uses a self-supervised learning module to augment the original dataset,constructs positive and negative sample pairs and differentiates samples by negative sampling method and relative similarity measure,then builds a graph neural network model based on knowledge graph and integrates multi-head attention mechanism to obtain entity features,and makes the model close to the real-world environment through dynamic scoring module,so as to improve the performance of the recommendation system and mitigate the impact of data sparsity problem.In this thesis,we conduct comparison experiments on Sports,Toys,and Yelp datasets to verify the effectiveness of the proposed model in this thesis based on hit rate,NDCG,and MRR in several dimensions.The experimental results demonstrate that the Self-KGNN model achieves hit rates of12.76% and 7.12% on the Toys and Yelp datasets,which is an improvement of about12.92% and 15.02% compared to the optimal hit rate of the comparison model,and the effectiveness of this method to solve the data sparsity problem is demonstrated by sparsity experiments.