The Research On Several Problems Of Collaborative Recommendation Based On Rating Prediction

With the rapid development of internet technology, more and more users purchase the different kinds of information and services they wanted through the internet. And the users are helpless facing the information overload problem. And the collaborative filtering technical become the widely used in various types to overcome the information overload. While collaborative filtering system has three main problems:inherent sparsity, scalability, cold start. Domestic and foreign scholars are working on several factors from this, with the purpose of improving the quality of the recommendation system. However, three main problems still need to be researched and solved, this study we aimed at these problems, and our main work is as follows:Firstly, this work analyzes the drawbacks of the similarity and prediction computational formulas in memory-based collaborative filtering algorithm, accordingly, proposes an improved collaborative filtering algorithm based on rating features, modified both computational formulas, and plenty experiments results on two different datasets prove that this improved algorithm can improve the accuracy of prediction, and solve the cold start problem in collaborative filtering system.Secondly, we analyze the influence accuracy of training parameters on final model to predict the ratings in matrix factorization model, meanwhile an automation training parameters determine algorithm is proposed. Extensive experiments on dataset validate this algorithm can select the optimal parameters to training the model in each iteration, meanwhile takes the shortest period of time to get the right training parameters for learning the prediction model.Besides, a comprehensive study is conducted to analyze why and how rounding affects the performance of collaborative filtering algorithms in terms of rating prediction accuracy as for the ratings in dataset is integer. And proving why rounding the prediction is necessary in post-processing of the predicted ratings and improving the performance and quality of the recommendation systems. Two new rounding approaches based on the predicted rating probability distribution are proposed. Extensive experiments on different data sets validate the correctness of our analysis and the effectiveness of our proposed rounding approaches.