Research On Interpretable Recommendation Methods For Sparse Data

Nowadays,people's daily behaviors and decisions are increasingly influenced by recommendation system in areas ranging from online shopping,audiovisual entertainment to lifestyle travel.The recommendation system means modeling based on the user's history and item-related attributes,inferring the user's interests from the huge amount of information,and helping the user find items of interest.The core technology of recommendation system is how to dig out effective information for users from the huge amount of information.However,the data sparsity problem in traditional recommendation algorithms can lead to poor recommendation quality and cannot improve the user's satisfaction with the system.In addition,the lack of explanation of the recommendation results weakens user's confidence in system.In order to solve such problems and improve the user experience of the recommendation system,this paper is developed from the problems of data sparsity and unexplainably of the recommendation system,improves the accuracy and interpretability of the recommendation systems,then introduces the knowledge graph into the recommendation system,and proposes the recommendation probability prediction model and the path-level recommendation interpretation model,respectively.As the first task of this paper,the accuracy of recommendation system is still the primary research point of the recommendation algorithm,and the data sparsity is still the bottleneck that restricts the further improvement of recommendation quality.To alleviate the problem of data sparsity and improve the accuracy of recommendation,relevant research works to analyze user and item information from different perspectives are carried out.However,the following shortcomings still exist:1)In the traditional collaborative filtering algorithms,the auxiliary information is integrated into the recommendation,and the relevance of the data cannot be captured,which limits the further improvement of recommendation accuracy.2)The deep collaborative filtering algorithm is characterized by data starvation and lack of a deep understanding of the data due to mechanical data fitting,which leads to deviations in the accuracy of recommendations:? We propose a KGE-CF model.This model starts with the user's interaction history item.It utilizes structured data,namely the knowledge graph between items,to mitigate the sparsity of item data information through semantic dependencies.Secondly,it combines the learned user preference and item features into the multi-layer perceptron network to fully learn the highlevel interaction information between users and items,to predict the user's preference probability for items.The experiment shows a good prediction effect.The interpretability of the recommendation systems is the second task of this paper.The interpretability of recommendation results is the key to improve system transparency,user trust,and satisfaction.As an important branch in the field of explainable artificial intelligence,the research of explainable recommendation systems has attracted the attention of finance,medical treatment,and other fields.The interpretability of the recommendation systems can further enhance the user's recognition and trust.It also provides an appropriate background for the development and research of explainable machine learning.To make the recommended content more acceptable to users and improve user experience,the main work done in this paper is as follows:? A KPCRN model is proposed in this paper.Firstly,we use deep search to extract relevant semantic paths of user history interaction items from the knowledge graph.Then,we use text convolutional neural network and recurrent neural network to model the connection path between the user and the item.Finally,we use weight pooling to get the user's preference probability for the item.The model displays the extracted path,so as to make the recommendation system has path-level interpretation.The experiments on real datasets(Movielens,Book-Crosings,Last.fm,and KKBox).The experimental results show that the high-order collaborative filtering recommendation model proposed in this paper can reduce data sparsity and improve the effectiveness of recommendations.The experiment also shows that it has a good prediction effect.The knowledge path awareness convolutional recurrent network model proposed by us can not only improve the system performance,but also effectively solve the problem that the recommendation result can not be interpreted,and provide path-level display and interpretation of the recommendation results.