Dot-product based preference preserved hashing for fast collaborative filtering

Recommendation is widely used to deal with information overloading by suggesting items based on historical information of users. One of the most popular recommendation techniques is matrix factorization (MF), in which the preferences of users are estimated by dot products of their real latent factors between users and items. Although MF can achieve high recommendation accuracy, it suffers from efficiency issues when making preferences ranking in real space. Hash retrieval technique can be applied to recommender systems to speed up preferences ranking. Due to the existence of discrete constraints in learning hash codes, it is possible to exploit a two-stage learning procedure according to most existing methods. This two-stage procedure consists of relaxed optimization by discarding discrete constraints and subsequent binary quantization. However, existing methods have not been able to well handle the change of dot product arising from quantization. To this end, we propose a dot-product based preference preserved hashing method, which quantizes both norm and cosine similarity in dot product respectively. We also design an algorithm to optimize the bit length for norm quantization. Based on the evaluation to several datasets, the proposed framework shows consistent superiority to the competing baselines even though only using shorter binary code.