Research On Explainable Personalized Recommendation Algorithm Based On Deep Learning

As a personalized information filtering tool,recommendation system has been more and more widely used with the advent of the era of big data.At the same time,recommendation systems are also facing development and challenges,such as data sparseness,lack of interpretability and other issues.The existing recommendation model is difficult to deal with sparse structured data,ignores the preference information contained in the user behavior sequence,and cannot provide a reasonable explanation for the recommendation result.In order to solve the above problems,this thesis focuses on the task of click-through rate prediction,deeply researches the recommendation algorithm based on deep learning,proposes two explainable personalized recommendation models,and designs and implements a financial product recommendation system.The main contributions of this thesis are as follows:1.This thesis proposes a recommendation model ESAt Int based on explicit sparse attention mechanism to solve the problem of sparse structured data and difficulty in modeling feature interaction.First,ESAt Int automatically maps high-dimensional sparse features to low-dimensional dense attention spaces,and by stacking interaction layers based on explicit sparse attention networks,feature interactions of a specific order are obtained.Then,ESAt Int can explain the recommendation results at the feature interaction level according to the different attention distributions of different feature interactions.Finally,the experimental results show that the algorithm's measurement index AUC has been improved and is explainable at the feature interaction level.2.This thesis proposes an explainable personalized recommendation algorithm DSKGAT based on the session graph attention network and knowledge enhancement.This method solves the problem that the sequence recommendation algorithm has limited expressive ability when capturing complex user preferences and it is difficult to obtain fine-grained user interest from interactive sequences.First,DS-KGAT divides the user sequence into multiple sessions to construct a session graph,and uses the session graph attention network to model the time evolution of user interest.At the same time,the builtin knowledge enhancement module spreads the user's conversational interest on the set of knowledge entities to form the user's multi-level preferences.Finally,DS-KGAT can provide a basis for the recommended explanation of the target item by tracing the user's interest transmission path.Experimental results show that the algorithm has good validity and explainable.3.In order to verify the feasibility of the explainable personalized recommendation algorithm proposed in this thesis,a financial product recommendation system based on ESAt Int and DS-KGAT is implemented.The system framework uses the Flask framework,the data storage uses the MYSQL database,and the user interacts with the user through the Web browser.The recommendation system can provide users with financial product recommendation services,and show users information such as financial product codes and issuance prices.At the same time,it also provides an explainable recommendation function,which lists the reasons for the recommendation while making the recommendation.