Learning To Hash For Collaborative Filters Ensemble

With the boom of web services,recommender systems play an increasingly important role in helping us to make use of all kinds of information effectively.Collaborative Filtering(CF)recommendation which can only use the history records of users to conduct a personalized recommendation fast is the mostly used technique now.However,the sharply increased scales of users and items within such web services make the task of recommendation more challenging than it used to be.Recently,the promising solution is to hash users and items in the form of binary codes,so that the recommendation can be efficiently made in Hamming space.Nowadays,most of the existing CF hashing methods have large encoding loss due to the oversimplified modeling on the continuous vector space and "two-stage" learning scheme.Thus,they usually need to utilize long codes to save more information,which may cause more extra costs and against our motivation for efficiency.In this paper,the deep researches and discussions on how to apply the hashing techniques into CF hashing methods are conducted and the specific contents are as follows:(1)This paper proposes a Binary Collaborative Filtering Ensemble(BCFE)method which ensembles users' and items' anchor approximation smoothness constraints on the foundation of the matrix factorization in the Hamming space.BCFE could preserve the original data geometry in the binary codes to some degree.As for the optimization loss of "two-stage" learning scheme,this paper devises a Discretization-like Bit-wise Gradient Descend(DBGD)solution which integrates the quantization stage into the optimization.(2)In order to save more information in the learned binary codes,this paper utilizes extra side information via auto-encoders to obtain more abstract features which can make contribution to the binary codes learning.(3)In order to maintain the specificity of users and items features,this paper adds extra individual feature smoothness constraint into BCFE.According to the limitation of DBGD,this paper further designs a discrete method to directly learn the shorter binary codes.Extensive experiments on three benchmarks validate the superiority of the proposed approaches in comparison to the state-of-the-art methods.