A personalized recommendation approach based on content similarity calculation in large-scale data

Abstract

Recommendation algorithms are widely used to discover interesting content for users from massive data in many fields. However, with more diversification of user requirements, the recommended accuracy and efficiency become a serious concern for improving user satisfaction degree. In this paper, we redefine the concept of content similarity by combining search words with personalized search references and describing their dimensions, then propose the calculation method of content similarity by defining the Hamming distance among current keywords, classified items and historical keywords. Through the pretreatment of support vector data description (SVDD), we may find specific tendency from the personal preference of classified items and present the final recommendation results arranged from high similarity to low one. Simulation experiments show that our proposed approach improves recommendation performance over the other two classical algorithms by an average of 17. 2% and reduces the MAE by 6. 3% on our large-scale dataset. At the same time, our proposed approach has a better performance on recall rate and coverage rate, and user satisfaction degree is also improved at higher extent.