Employing Spectral Domain Features for Efficient Collaborative Filtering (1703.01093v1)

Abstract: Collaborative filtering (CF) is a powerful recommender system that generates a list of recommended items for an active user based on the ratings of similar users. This paper presents a novel approach to CF by first finding the set of users similar to the active user by adopting self-organizing maps (SOM), followed by k-means clustering. Then, the ratings for each item in the cluster closest to the active user are mapped to the frequency domain using the Discrete Fourier Transform (DFT). The power spectra of the mapped ratings are generated, and a new similarity measure based on the coherence of these power spectra is calculated. The proposed similarity measure is more time efficient than current state-of-the-art measures. Moreover, it can capture the global similarity between the profiles of users. Experimental results show that the proposed approach overcomes the major problems in existing CF algorithms as follows: First, it mitigates the scalability problem by creating clusters of similar users and applying the time-efficient similarity measure. Second, its frequency-based similarity measure is less sensitive to sparsity problems because the DFT performs efficiently even with sparse data. Third, it outperforms standard similarity measures in terms of accuracy.