Research On Data Quantity And Quality Sensitive Recommender Systems

With the rapid growth of online data,the information overload problem is getting worse and it becomes more difficult for users to find information meet their demands.Recommender system has been proposed to ease this problem.It helps users to find the potential valuable and personalized information via mining user preferences from their historical behaviors.In the last decades,recommender system has gained great attention from academia and industry.It has been widely applied to different applications,includ-ing the E-commerce,social networking,online entertainment,online reading and learn-ing.In current days,the related recommendation technique is evolving and researchers still attempt to design more effective,efficient and unified algorithms.Current recommendation algorithms usually assume that the distribution of data is proper and user behave consistently and ignore the difference of quantity and quality in user data.Then they use all the data to train model with simple processing or without pre-processing.However,user behavior data show differences in quantity and quality in real applications.In real systems,some users tend to produce little data while others generate rich data.At the same time,users face noisy ratings.In the age of big data,user data show imbalanced distribution and inconsistence.These issues will greatly influence the performance and efficiency of algorithms,thus research on the data quantity and quality sensitive recommender system has a very important research and application value.This thesis focuses on the data quantity and quality sensitive recommender system,and the main contribution are listed as follows.1)A data quantity sensitive recommendation methodTo deal with the quantity difference issue of user behavior data,we investigate the relationship between user data quantity and the performance of different algorithms.We show that it is not necessary to use all collected for model training,especially for active users.Then we study the needed quantity of user data for recommenaiton algorithms in terms of machine learning process.Finally,we design different sampling strategies.The experimental results show that this method improve recommendation at lower cost.2)A group transfer learning recommendation method based on user coherence.To deal with the quantity difference issue of user data,we propose a group learning recommendation method based on user coherence.We first introduce the concept of user coherence and group users based on it.And then apply different processing methods and recommendation algorithms for various groups.Finally,we transfer learning from high quality user groups to low quality user groups.The experimental results demonstrate that the proposed method significantly improve the recommendation of low quality user groups and improves the overall accuracy of recommendations.3)A semantic enhanced Bayesian Personalized Ranking with the semantically comparable item pairsBPR(Bayesian Personalized Ranking)is the popular method for the one class collaborative filtering problem and it has been widely used for implicit feedback recom-mendations.However,it suffers from the noisy data and tends to select low quality item pairs for model learning and leads to slow convergence and slow computation.To cope with these problems,we propose a semantic enhanced Bayesian Personalized Ranking with the semantically comparable item pairs.We introduce semantically comparable item pairs and improve the item pairs for model learning.The experimental results show that the proposed method could lead to a speeded and stable model learning with less semantically comparable item pairs.4)A recommendation framework of transfer learning user groups partitioned based on data quantity and quality.Current recommendation algorithms typically ignore the difference of data quantity and quality.We propose a recommendation framework of transfer learning user groups partitioned based on data quantity and quality.This framework includes measure of data quantity and quality,user grouping,data sampling,noisy data processing,and algorithm selection and transfer learning of user groups.It improves the performance of different algorithms designed for the rating prediction and TopN recommendation tasks.