Combining Multi-source And Heterogeneous Data In Recommender Models And Systems

In the information age,data matters a lot.The information overload is just caused by massive data.It is difficult for people to analyze and process information,so as to find the content that they are interested in.Information producers try to make their information attractive to control the traffic and generate revenue for the platform.Rec-ommender system,a kind of information filtering system,has greatly alleviated the contradiction.It serves for people by mining user behaviors,modeling users'interests,and predicting their preferences for potential information.The research of recommender system first starts from information filtering.Since then,with the development of information networks,the change of user behavior and the rise of computer science such as machine learning,various recommendation models and algorithms have been delivered.Collaborative filtering is the most popular one to be researched and implemented in the industry.It learns hidden features from prefer-ence interactions between users and items,and then predicts the users' favorite items.In real applications,the relations of users and items are very sparse.In addition,the new users and items do not have effective features,which fiercely degrades the performance of recommender systems.At present,the main solution to the problem is integrating re-lated auxiliary data,including user social network information,user's review text,user tags,and other feedbacks,etc.These sources of data are diverse and have different structures.The difficulty now comes to design a synthetic mechanism,where the data can be combined properly.The final goal is to improve the recommendation accuracy.In terms of problems above,this thesis mainly focuses on hybrid recommender systems to combine multi-source and heterogeneous data in a unified recommendation model.The main contributions include:1.This thesis proposes a hybrid recommendation model using rule-based machine learning techniques to combine multi-source of heterogeneous data.By parsing the auxiliary data like tags,text,social relations,and other data related to users and items,the system has some regulations and patterns to tune the results given by col-laborative filtering and content-based methods.In real code community forum and financial news recommender systems,more accurate and reasonable recommenda-tions can be generated to improve user satisfaction.2.The thesis proposes a multi-task recommendation framework,MultiCombine,to fuse multi-source heterogeneous data.With deep learning technology,the model mines effective latent feature from related auxiliary data.The common features of users are shared by the multi-tasks while the private factors are associated with specific tasks.By this means,the common knowledge of users can be transferred.Experiments have proved that the fusion model can finally promote the performance of downstream tasks.The combination of deep representation learning and neural networks will construct user profiles with scalability and interpretability,and make it flexible and adaptive to multiple tasks.3.Based on the synthetic recommendation models and algorithms mentioned above,the thesis designs and implements a personalized financial news recommender sys-tem which combines multi-source data of users.The system makes different recom-mendations for different users to meet their needs.It does not generate a same list of candidates ranked by time and popularity.The system will consider the reading history as well as affiliated information like tags and social relations of users.This kind of approach is more reasonable than traditional ones.This thesis focuses on the mechanism of combing multi-source and heterogeneous data in recommender systems.The study starts from the rule-based approaches to ma-chine learning,especially deep learning techniques.Experiments in certain scenarios and datasets prove that the fusion of multi-source heterogeneous data is an effective way to improve the performance of recommender systems and mitigate the data sparsity and cold start problems.