Study On Context-based Recommendation

In the era of massive data,users’ online discovery of meaningful content is increasingly influenced by recommendation systems.Recommendation systems can combine the user’s own characteristics and the environment to make targeted recommendations to the user.How to better model the contextual information in the recommendation system becomes a key factor to improve the performance of the recommendation system.How to use contextual information well is the key to describe the recommendation scenario,and the diversity and complexity of contextual information also present important challenges for the research of recommendation systems based on contextual information:(1)user ratings,as user-initiated contextual information,are often used as the ground truth to express users’ likes and dislikes,but there are often inconsistencies between user comments and user ratings;(2)traditional recommendation datasets rarely record negative feedback of users,and traditional recommendation models usually only model positive feedback information of users,which cannot accurately capture information that users do not like;(3)during the interaction between users and recommendation systems,users’ recommendation lists will dynamically change,which increases the difficulty of capturing users’ preferences;(4)most of the existing recommendation datasets contain only partial contextual information and are small in scale,lacking of large scale recommendation datasets with rich contextual information.In response to the above challenges,this thesis carried out research on recommendation methods based on contextual information:(1)we did a crowdsourcing annotation analysis of rating unreliability,designed a review-based rating revision algorithm,and carried out validation of revised ratings;(2)we tried to model user negative feedback directly and proposed a negative feedback aware hybrid sequence recommendation model,and the proposed method improved the area under the curve(AUC)on the Xing dataset by 2.98%;(3)we applied recurrent neural networks to model the user dynamic click sequences,and explored how to improve user satisfaction by portraying the dynamic changes of user diversity needs through the dynamic changes of user recommendation list;(4)we constructed Zhihu Rec dataset,which based on Zhihu knowledge-sharing Q&A community,as well as containing user query keywords.This thesis focuses on the effective use of contextual information to improve the recommendation method,and the main innovations are as follows:(1)we propose the problem of inconsistency between user comments and user ratings,propose a rating prediction model based on the content of comments,and the ratings after revision by the model can effectively improve the effectiveness of the rating prediction task;(2)we propose a negative feedback-aware hybrid sequence recommendation model,the model uses contextual information more fully,and the model can effectively use the information that users do not like to improve the recommendation accuracy;(3)we propose a recommendation model that combines list diversity,user click history diversity and sequence information,the model better uses the dynamically changing contextual information;(4)to the best of our knowledge,our Zhihu Rec dataset is the largest recommendation dataset with the most comprehensive contextual information.