Modified Similarity Algorithm for Collaborative Filtering

Abstract

Collaborative filtering (CF) is one of the most applied techniques in recommendation systems and has been widely used in various conditions. The accuracy of the CF method requires further improvement despite the method’s advancement. Numerous issues exist in traditional CF recommendation, such as data scarcities, cold start and scalability problems. Since the data’s sparsity, the nearest neighbors formed around the target user would cause the loss of the information. When the new recommended system started, the information about evaluating is poor, the result of recommend is also poor. About the scalability problems, under the background of big data, the complexity and accuracy of calculation is facing a great challenge. Global information has not been fully used in traditional CF methods. The cosine similarity algorithm (CSA) uses only the local information of the ratings, which may result in an inaccurate similarity and even affect the target user’s predicted rating. To solve this problem, a modified similarity algorithm is proposed to provide high accuracy recommendations, and an adjustment factor is added to the traditional CSA. Finally, a series of experiments are performed to validate the effectiveness of the proposed method. Results show that the recommendation precision is better than those of traditional CF algorithms.