Cross-modal Hashing For Large-scale Recommender Systems

With the advent of the “big data” era,very large amounts of data have accumulated on the Internet.Nowadays,in many online web services,it is a vital task to precisely recommend relevant items from massive candidates.This has attracted widespread attention among experts and scholars.Under this circumstance,a large number of recommendation algorithms has emerged.Collaborative filtering based on latent factors is one of the popular methods among them.However,as the amount of data continues to grow at an alarming rate,the traditional collaborative filtering methods are too expensive to recommend.To confr ont with the scalability issue,a discrete latent factors learning based recommendation method which inspired by hashing is advocated.This kind of method permits exact top-K item recommendation with sub-linear time complexity and greatly improves the efficiency.But there are also the following limitations:(1)The existing discrete methods only consider the cross-view user-item relations,that is,the user-item ratings matrix,to learn user/item latent factors.However,these methods overlook the intra-view user-user/item-item affinities.(2)The existing discrete methods face the notorious cold-start problem and long-tail problem of the recommender systems.The cold start problem means that the recommender systems cannot accurately recommend new users/items due to the lack of corresponding rating data.The long-tail problem means that the recommender systems tend to recommend popular items,while overlooking some niche items that may be favored by certain users.In response to these two issues,the main work and innovations of this dissertation include the following three points.(1)Discrete manifold-regularized collaborative filtering:Inspired by manifold learning,we propose a discrete manifold-regularized collaborative filtering method which jointly exploits cross-view user-item relations and intra-view user-user/item-item affinities to learn the compact binary codes of users/items in the hamming space.To reduce the quantization error,the proposed method use the discrete cyclic coordinate descent algorithm to directly optimize the discrete objective function.Experiments demonstrate the proposed method not only outperforms the state-of-art discrete latent factors learning methods but also can effectively improve the performance of cold-start recommendation by transferring knowledge from the old users/items to the new ones.(2)Discrete collaborative filtering based on low-rank and sparse decomposition:On the basis of the above method,in order to further improve the item diversity of recommendation algorithm,we propose a discrete collaborative filtering method based on low-rank and sparse decomposition.This method decomposes the user-item ratings matrix into two parts: ratings matrix for popular items and ratings matrix for niche items,and then imposes the low rank constraint and the sparse constraint to the two matrix respectively.Discrete latent factors are learned under this framework.Similarly,the proposed method also develop an alternative discrete optimization algorithm to reduce the quantization error.The experiment results show that the proposed method is not only superior to the state-of-art discrete latent factors learning methods,but also can effectively improve the performance of long-tail recommendation.(3)Discrete collaborative filtering based on Siamese graph convolutional networks:To further mine the intra-view user-user/item-item affinities,we propose a novel discrete collaborative filtering method based on Siamese graph convolutional networks.The proposed method construct user-user/item-item graph from user-user/item-item affinities respectively.The vertices of graph represent users/items and the edges represent similarities between vertices.With the help of graph convolutional networks,the proposed method can extract user/item featu res and learn user/item binary codes from the user graph and the item graph simultaneously.Experiments demonstrate the proposed method outperforms the existing discrete latent factors learning methods and has a lower training cost.