Research On Hybrid Recommendation Algorithm And Its Parallel Implementation Based On Regression Strategy

With the rise of the Internet and mobile intelligent terminals,humans are facing with serious information overload.In order to solve this problem,the recommender system came into being.The recommender system establishes the user model by analyzing the user's preference and behavior information,and actively recommends the information that the user may be interested in.However,due to the rapid growth of data size,the performance of the proposed algorithms decreases drastically.In view of the dilemma faced by the current recommender system,this paper puts forward corresponding improvement measures.The main research results and innovation on this dissertation are listed as follows:(1)In order to solve the data sparsity,this paper obtains the item feature vector through the latent factor model which was commonly used in the recommendation system;Then,the linear regression and support vector regression are introduced to obtain the user's interest on latent factor.Finally two kinds of algorithm models are obtained,Which are called local weighted linear regression based on latent factor model(LR-LFM)and support vector regression based on latent factor model(SVR-LFM).The former is suitable for situations where real-time and model availability are required,while the latter is suitable for systems which pursuit algorithm precision in data sparsity.(2)To address scalability and real-time issues,we choosed Spark which is more suitable for iterative computation to implement the parallel form of the proposed SVR-LFM algorithm.We analyzed the data dependency and communication overhead of the algorithm,finally gave a detailed implementation flow of parallelization.(3)At last,we used python to implement the algorithms mentioned above,trained the corresponding parameters through the MovieLens dataset,verified the performance of the algorithms.Experiments showed that the algorithms proposed in this paper has improved the accuracy significantly in sparse data and the parallel form of SVR-LFM algorithm has a nearly linear speedup.