Research On Recommendation Technology Based On Item Implicit Information

The recommendation system is a tool that can meet the individual needs of users.It can effectively alleviate the problem of information overload on the Internet.Collabo-rative filtering based matrix factorization is the most widely used technique to solve the recommendation system rating prediction problem.At present,a large amount of exist-ing work focuses on the advanced optimization method based on matrix factorization.In order to alleviate the data sparsity and cold start problems,many researchers added additional information to the corresponding optimization task,such as user social in-formation,item attributes,user reviews,and geographic locations,etc.However,most of the existing work directly integrates these explicit information into the optimization goals.Few jobs focus on the further exploration of explicit ratings information.Rat-ings,as an explicit representation of user and item interactions,contains rich implicit information.These implicit information can be used as a useful supplement for the recommendation system to understand the user's preference process,and they improve the performance of the recommendation system.This thesis aims to mine implicit information from the rating data to further im-prove the performance of the recommendation system.On the one hand,the impact of the popularity of each item is explored from a global perspective.On the other hand,the implicit relationship between two items is tapped from a local perspective.We mix above two influence factors into the traditional matrix factorization model and propose a recommendation model based on the implicit information of the items.In addition,We combine recommendation model based on the implicit information and recommen-dation model based on social information into a more effective model.The purpose of this thesis is to solve the problem of rating prediction in the rec-ommendation system.By analyzing the implicit information behind the ratings,we introduce two influence factors of the item and design a recommendation model based on the item information.Then we try to ensemble the model with the recommendation technology based on the user's social information.The work of this thesis includes:1.We propose a recommendation model based on item implicit information.The pre-vious recommendation models based on item information mostly only added addi-tional data sources directly,but they neglected to mine the item implicit informa-tion in ratings.We propose two method to mine implicit information of items from global and local perspectives.These implicit information can further characterize the features of items.We integrate these implicit information into a traditional ma-trix factorization model in a reasonable way.2.We propose a recommendation framework for integrating multiple information.We expand on the recommendation model based on the implicit information of the items,and combine the social recommendation model into a multi-information en-semble recommendation framework.We integrate two mainstream social recom-mendation models into the framework and verify versatility and effectiveness of it through real experimental data.The framework not only has a unified optimization goal,but also further enhances the performance of the recommendation model.3.We implement a simple prototype recommendation system based on the proposed multi-information recommendation framework.The system mainly includes sev-eral modules:data acquisition,data preprocessing,model training and model rec-ommendation.The basic flow of the model is shown,which reflects the rationality of the proposed multi-information recommendation model to some extent.