Hybrid Latent Vector Model-and Item-based Collaborative Filtering Recommendation

Along with the Internet having achieved widespread successes in the daily life, the problem is how to find the most significative data we need quickly, while facing the mass data with lots of useless informations on the web. In order to overcome the problem called information overload, personalized recommendation system has been emphasized. They study the user behaviors, contents of the site and etc, and use a large class of machine learning alogorithms including knowledge discovery and filtering algorithm to pick out the meaningful data which the people may like. As the web2.0 rapidly develops and the concept of web3.0 raised, people are getting more and more taking part in the update processes of the web site. They design the internet products and deliver them by uploading them on the website. Also they may give some feedbacks about the products, and share them with friends. It is important to collect the effective informations about the feedback from users, and uses the knowledges about statistics and pattern recognition to model the mathematical model. Taking advantage of computing capability by the computer, the recommendation system analyse the user preferences, and make the applications work intelligentized to think out the users's ideas automatically.Firstly, the knowledge background of the personalized recommendation is introduced. We also introduce the recommendation's architecture, and set forth its advantages. Secondly, as the algorithm being the core of the system, thereinafter the paper focuses on it mostly. Collaborative filtering(CF) is one of the most popular techniques that help users to make choices and find relevant items in a recommend system. And it is scaleable conveniently, and has good accuracy of the recommendation's results. Collective intelligence and its implementation can develop a collective of users preferences. Collaborative filtering is devided into two parts: Memory based and Model based. Memory based algorithm includes user-based and item-based, and it supplyies good explanation in theory. Latent Semantic Analysis does not need extraneous informations, and construct the space of latent semantic on the rating matrix to find the latent interests of users.The third part introduces the two kinds of the nearest neighbor algorithms: use the distance formula to compute the similarity bewteen the users(items) and use the minimum mean squared error to regress the correlation coefficent. Also the algorithms of computing the similarity and neighbor selections had been optimized. The forth part combines the gaussian probabilistic latent semantic analysis and the improved item-based collaborative filtering together. That method models a mixture of user communities and generates similar items of target item respectively. The fifth part introduces user bais into the regularization SVD model, and uses the gradient descent method on feature-based, and combine the neighbor algorithms with dynamic parameters of hybrid recommendation . Experiments on the movielens dataset show that the proposed approachs compares favorably with other collaborative filtering techniques.