Collaborative Filtering Recommendation Algorithm Based on Two Stages of Similarity Learning and Its Optimization

Abstract In order to improve the method of similarity measure in the existed collaborative filtering recommender system, a collaborative filtering recommendation algorithm, based on two stages of similarity learning, was proposed. It aimed at improving the prediction accuracy of the algorithm through few iterative calculations. The algorithm achieved a higher accuracy, which mainly attributed to the combination of the reduced gradient method and the K-nearest neighbor algorithm. Similarity support and solution to over-fitting were also adopted to make the algorithm more superior. The experimental results show that the collaborative filtering recommendation algorithm based on two stages of similarity learning, on some conditions, not only outperforms the existed method in terms of the error performance but also makes dynamic similarity adjustment possible.