Research On Recommendation System Based On Causal Embedding

To alleviate the problem of information overload caused by the explosive growth of data,recommendation systems have emerged and been widely applied in various fields.Traditional recommendation methods usually use collaborative filtering,content filtering,or hybrid methods to recommend products or services to users based on associations,without considering causal relationships between users and items,which may lead to unreasonable or low-quality recommendations.Compared with traditional methods,causal recommendation methods have stronger interpretability.Causal recommendation systems can help users better understand why they receive certain recommendations,thereby enhancing their trust and satisfaction.In addition,causal recommendation systems can effectively handle problems such as observed data bias,multi-factor influence,and consideration of causal relationships,making their recommendations more reliable.However,causal recommendation methods still face some challenges,such as ignoring complex relationships between items in causal embedding,leading to performance degradation,and using loss functions that do not consider similarity between items,resulting in lower generalizability of the model.To address these problems and further improve the performance and scalability of causal recommendation systems,this paper proposes the following solutions:(1)Use causal embedding to separate personal interests from conformist behavior and embed them into the model,while adding dependencies between items.First,use Graph Convolutional Networks(GCN)to encode and embed complex relationships between items.Second,introduce similarity or dependency weights between items in the loss function of Bayesian personalized ranking(BPR),and propose a new debiasing framework called XDICE,which has the ability to consider complex relationships between items.(2)Use the new debiasing framework XDICE proposed in(1)to improve the Neural Graph Collaborative Filtering(NGCF)model based on GCN.Replace the embedding propagation layer in the NGCF structure with a causal embedding propagation layer,combine the recommendation model with the debiasing framework,and use the weighted Bayesian Personalized Ranking(wBPR)loss function to maximize the difference between scores of popular and unpopular items,which improves the model’s ability to alleviate data and popularity bias.(3)Design and implement a web-based movie recommendation system,and integrate the improved NGCF recommendation model from(2)into the system to meet users’ convenient and fast scenario experience needs.Experiments were conducted using the Movielens-10M and Netflix Prize datasets,and objective evaluation metrics such as Recall,Hit Rate(HR),and Normalized Discounted Cumulative Gain(NDCG)were used to evaluate the recommendation models.The XDICE debiasing framework proposed in this paper outperformed the DICE debiasing framework on all three objective evaluation metrics.