Key Technologies And Implementation Strategies Of Recommendation Systems Based On Incremental SVD And RBM

With the coming of information society, information related industry got rapid development. This brings the rapid expansion of information, as a result, information overload, under the background recommender systems have been proposed. Recommender system is a technology that can recommend useful items to the user.In recent years, recommender system has become a hot research field, thus it attracts many scholars, not only in the industry, but also in academia, besides,many competitions, such as Netflix and related international conferences held every year, make recommender system become an independent discipline. Although the technology of recommender system is becoming more and more mature with the constant development of new technologies and new methods, there are still some problems to be handle, for new users, new items. According to the data set, several typical data preprocessing techniques are used, for example, in order to make the model more accurate, de-noising method, sampling with replacement method and filling the missing value with average value method are adopted; in order to improve the robustness of the model, artificial noise is added, etc..The model is easy to overfitting and other issues.In this paper, someimprovement methods for theexisting problems have been proposed, described as follows:(1) Although the recommender system based on singular value decomposition(SVD) technique is better than the traditional recommender technology, the computational cost is high, both the time complexity and space complexity, and the shortcomings is more obvious when data set is large. Therefore, based on SVD method, an incremental SVD building model is proposed, which using the folding-in techniques to make new users to a linear mapping. The proposed method can not only improve the scalability of the recommender system, but also greatly reduce the whole process(training and predicting) time complexity and guarantee the accuracy of the prediction, which is suitable for offline and online calculation of recommender system.(2) Recommender system based on the bipartitenetworkis a personalized recommender based on the weight method. The bipartite network directly compressed into a single mode mapping. Although the single mode has less information than the bipartitenetwork, but to preserve the original information by using the weight method, while taking advantage of the network dynamic resource allocation idea, to extract the information from the network directly by using the weight method, and the performance was significantly better than the global ranking method.(3) A recommender system based on RBM model is suitable for large scale data set, and the two layer undirected graph model is suitable for processing tabular data. It can make learning and reference effectively, and can combine the probability model and graph model to establish the connection between user and rating. The prediction results are more interpretable, and the performance is better than SVD method.Of course, with the increasing number of Internet users, the user's demand is also increasing, more various and more varied. However, the data sparse problem, feature extraction, cold start, over fitting and many problems still exist. Therefore, we need to continually improve the original recommender technology, and develop new technology to meet user needsto provide users with more convenient and effective user experience. Making the recommender system technology matures to play a more significant role in the industry.