Study On Recommendation Method Based On Actively Semi-supervised Collaborative Filtering

With the development of information technology(especially Internet technology),people are facing more and more serious problems of information overload while enjoying the convenience of information acquisition.In this context,personalized recommendation system as an effective technical means to solve information overload has emerged,and has become an indispensable part of people's life.Collaborative filtering(CF)is one of the core technologies for building recommendation system,which has been highly valued by academia and industry in recent years,but its performance is severely constrained by the problem of data sparsity.In view of this,an active semi-supervised collaborative filtering model for rating prediction task is proposed.The model uses both labeled and unlabeled samples to learn user preferences to alleviate the problem of data sparsity.In the framework of collaborative training,the model uses two neighborhood based CF algorithms to construct the basic recommendation.Each recommender independently predicts the unlabeled samples,and fills the samples with high confidence of its prediction into the other training set for the next iteration of the model.By repeating this process,the two recommenders can promote each other through sample exchange and continue to develop in the direction of performance optimization.In the process of iteration,the basic recommender adopts the active learning idea to select the unlabeled samples which have positive effects on the model's prediction performance for marking in order to accelerate the convergence of the model.At the same time,a confidence prediction strategy is designed to ensure that the pseudo-marking samples filled into the training set will not degrade the performance of the model.In addition,we have designed two kinds of methods to model the efficiency optimization,respectively is incremental similarity calculation method and the neighborhood range optimization method.The former is to separate the changing part from the overall calculation from the original similarity calculation formula,and the latter is to narrow the range of nearest neighbor search.These methods reduce the time cost of the model from different perspectives.The experimental results on three data sets show that the unlabeled data can effectively alleviate the problem of data sparsity,and the proposed model achieves better recommendation performance than other similar methods at the cost of a small amount of extra computational overhead.