A Research On Prior Knowledge Assisted Top-N Recommendations

With the development of Internet technology,e-commerce,cloud computing,mobile computing,and other related technologies,the era of Big Data is coming.As a result,the information overload problem makes it hard for people to access their preferred information.Fortunately,recommender systems(RSs)have been widely explored to alleviate this issue.Many scenarios,such as online shopping,news delivery,and community service,have embraced RSs to improve user experience and platform value.Thus,RSs become one of the critical applications of artificial intelligence.They attempt to provide the top-N items that meet peoples’ expectations based on learning historical data.However,existing approaches utilize prior knowledge inadequately,which could limit the performance of RSs.To this end,this paper proposes several methods that can directly or indirectly utilize prior knowledge from the perspectives of feature engineering,model structure designing,and objective function optimizing.The purpose of this paper is to provide people with a more accurate top-N recommendation list,and the main contributions are as follows:First,the paper proposes a recommendation framework called 6)NNDnn,which can directly fuse a kind of prior knowledge.This framework combines rule-based and neighborhood-based approaches.At the same time,it considers the relations between recommendation sub-tasks by proposing a multi-task cooperation module for feature engineering.Besides,there is a weight learning mechanism based on genetic algorithms for different features.The mechanism pays different penalties for different classification errors to improve the effectiveness of recommendations.6)NND-nn has been applied in real-world medical treatment plan recommendations successfully.Second,to design a better neural network structure for top-N RSs,this paper proposes KA-Mem NN,which can model the user’s sequential decision-making process by considering the difference between user intention and user preference.The structure of KA-Mem NN well fits the prior knowledge that users have hierarchical intents.It helps the model to learn a more refined user representation.Furthermore,by considering the prior knowledge that users’ groups may be changing over time,this paper designs a neural network model called LSUG,which can dynamically identify users’ groups and evaluate each group’s influence on users’ decision-making processes.Both KA-Mem NN and LSUG can achieve a better recommendation performance.Third,since different optimization functions focus on different rules behind data,and some of them have inherent relations,this paper proposes CPL framework to combine pointwise and learning to rank(L2R)approaches based on multi-task ensemble learning.Then,the estimated value of each example is supposed to be approaching the real value and keeping the preferential relations between items.Moreover,to overcome the shortage of traditional ranking approaches that only aim at single items,this paper directly ranks the collection of items and proposed GTRM.The approach has a setbased optimization function,which is motivated by the prior knowledge that similar items should have similar positions in a recommended list.Thus,the model can improve the position of the visited items as well as the positions of the items in the same set that have not been visited yet.The designed optimization functions of both CPL and GTRM make the training process more reasonable.Based on multiple real-world datasets,this paper conducts comprehensive experiments to evaluate all proposed approaches.The results prove that the prior knowledge assisted strategy can improve the performance of the top-N recommendations.