Context-aware Personal Recommender System

Users are facing a predicament called“Big Data, Thin Knowledge” in the era of big data. Information recommendation(IR) technology can build user preference model with user historical interactions information, then recommend the information which is user’s potential interest to and alleviate the information overload problem in a certain extent. However, the traditional IR algorithms only consider the interaction between the user and the recommended object, few attention to the different environments’(Context) affect to users’ preference, which must be consider in the actual recommender system.Form the view of the above-mentioned facts, this thesis introduced the concept of context built on traditional IR algorithm. When build user preference model, take account of the different context’s effect. And two context-aware recommender algorithms were proposed. Meanwhile, in order to solve the problem of contextual information sparsity, we present the Localized Matrix Factorization(LMF)recommender algorithm based on contextual information entropy, which factor Block Diagonal Form(BDF) matrices.The main contributions are as follows:First, this thesis designed the scheme of individual IR system based on summarized the related studies. particularly introduced user modeling 、recommend object modeling and recommendation algorithm modules, then general process and design framework of information recommendation were given. In this sense,implemented design of context awareness recommender system by introducing the context information.Secondly, according to the IR system framework, this thesis proposed two context awareness recommendation algorithms:(1) Pre-filtering an IR algorithms.Context information filter by clustering technology to make rating data of cluster User-Item with similar context. Then, finding the highest similarity of cluster through perform recommendation algorithm by user collaborative filtering in cluster for object user.(2) post-filter context information algorithm. Without context information, by the traditional user collaborative filtering obtain an initial list of recommended.Combine the calculate probability of user in a specific context situation prefer certain types of projects and the result of collaborative recommendation, rerank the list. The experimental results show that compared with basic algorithm, both methods can effectively improve the performance of the recommendation algorithm.Then, to solve the data sparsity problem in recommender system, we present the LMF recommender algorithm based on contextual information entropy, which attempts to meet scalability by factorizing BDF matrices. In this algorithm treatedoriginal score matrix as a bigraph, image segmentation algorithm is iteratively applied to the bigraph community division transform original matrix into Recursive Bordered Block Diagonal Form(RBBDF). By mending sub-matrix’s reserved judgement conditions, adding consider contextual information entropy strategy on the basis of take matrix mean density into account. Then, using the factorization result on the diagonal blocks predict the zero blocks of BDF. Conversion to the ranks to score predicts matrix for collaborative filtering at last. Experiments prove that this method can effectively improve the performance of the recommendation algorithm.Finally, the effectiveness of algorithm proposed in this paper had been verified at LDOS-CoMoDa and MovieLens dataset. The experimental results indicate that in accuracy and root mean square error(RSME) has more obvious improvement.