Research On Explainable Recommendation Algorithm Based On Features And Reviews

With the development of e-commerce websites and various online information platforms,people can find items and information they are interested in through recommendation systems.And recommendation algorithms,as the core of recommendation systems,have been greatly sought in academia and industry in recent years.The traditional personalized recommendation algorithm is based on the interaction information between users and items,but the number and types of items in the system are growing and changing all the time,and cold start becomes a major problem for the recommendation system.With the deepening of deep learning in the field of recommendation,recommendation algorithms have improved in accuracy and scalability,but deep learning algorithms have the unexplainability of black box models,and the generated recommendation results often do not provide users with convincing explanations,and the lack of transparency and trust puts both recommendation systems and users in a difficult situation.Based on the above reasons,this paper focuses on the accuracy and interpretability of recommendation algorithms and does the following:(1)To address the cold start problem in recommender systems,XGBoost regression trees are constructed using objective item attribute information,and the task of predicting item ratings is completed by inputting item attribute information into the XGBoost model for a series of supervised training,and the cross-features of item are extracted from the model while generating the tree model as the interpretation of recommendation results.It also uses K-means clustering to perform coarse-grained partitioning of users and goods to save subsequent XGBoost training time.And a large number of experiments and analyses have been carried out on the Amazon dataset,and the results show that the proposed algorithm in this paper not only improves the recommendation effect,but also extracts the important attributes of the goods as the recommendation explanation.In addition,the introduction of K-means saves the time of model training and reduces the computational cost.(2)Aiming at the interpretive problem of recommendation system,a review based interpretable recommendation algorithm is proposed.The algorithm consists of two independent modules: the recommendation model for rating prediction and the text generation model for generating explanatory sentences.In the scoring prediction module,the convolution neural network is used to extract the features of the user and item comment information,and the weight is automatically assigned to the comment through the double-layer attention network.The high weight reviews are retained to prepare for the generation of interpretation sentences.Finally,the scoring prediction is carried out by fusing the features of the user and item through LFM.In the text generation module,the double-layer GRU network in HSS model is used to train the high weight reviews retained in the previous step to generate complete and explanatory sentences.Experiments show that the method proposed in this paper not only has good performance in rating prediction,but also generates complete explanatory sentences and processes,which can contain important characteristic words of users and items,and has good explanatory significance.