The Research On User Profiling And Time Awareness-based Hybrid Approaches For Recommendation Systems

Most of the users on the Internet face the information overload problem nowadays.Various recommendation systems and approaches are applied to address the information overload problem.Recommendation systems are generally used to provide users with relevant information or predict users' assessment of unknown items.Hybrid recommendation systems integrate different recommendation approaches and techniques to utilize their advantages,overcome their limitations,achieve better results,and increase the quality of recommendations and predictions.However,although the rapid development and considerable attention from both industry and academia,recommendation systems still face numerous challenges and limitations,such as the coldstart,data sparsity,user modeling and user preferences dynamic.Many academic studies only concentrate on user-to-user or item-to-user similarity algorithms with no regard to user profiles and their efficiency and weights.Besides,insufficient attention is put on the dynamics of users' preferences.Consequently,apart from increasing the accuracy of recommendation outcomes,the efficiency and effectiveness of recommendations could be further improved by addressing the challenges and limitations of recommendation systems.Therefore,this thesis mainly focuses on user modeling and the hybrid recommendations approaches.More specifically,it studies the different approaches of user modeling to address the challenges and alleviate the limitations of recommendation systems.Moreover,user-generated content such as users' reviews is also utilized for user modeling to improve user profiles and enhance the recommendations' performance.Furthermore,a new hybrid recommendation approach for individual users to address the user preferences changes is proposed.The proposed hybrid recommendation approach covers content-based and neural collaborative filtering,extended users' and items1 information for user modeling,and top n recommendations and rating prediction tasks.Additionally,the extended users' and items' information is also used in knowledge graphs to utilize the relationships among users,items,attributes,and features in order to improve the rating prediction performance.The performance of the proposed approach's is thoroughly evaluated using a real-world dataset and compared with several baseline recommendations algorithms and approaches.Various evaluation metrics,such as precision,recall,novelty,diversity,MAE(Mean Absolute Error),and RMSE(Root Mean Square Error)are applied to evaluate the proposed approach thoroughly and comprehensively in the experiments.The experimental results show that the proposed approach outperforms the baseline algorithms in top n recommendations and rating predictions tasks and better addresses the user preference changes issues.Furthermore,this study also designs and implements a recommendation system prototype,relying on the proposed hybrid recommendation approach.Such a prototype provides the graphical user interface of the recommendation system's administrative panel,which allows viewing,executing,and controlling a variety of functions of the recommendation system.The main functionality includes view and search of all users and items in the system,build users' profiles for each user,construct and view different knowledge graphs,provide top n recommendations and generate rating predictions for each user.This study contributes to academia in the research of user modeling,hybrid recommendations development and evaluation of recommendation systems.Moreover,it contributes to the industry by designing and developing a recommendation system prototype based on the proposed hybrid recommendation approach.