Research On Hash Recommendation Methods Based On Matrix Factorization

With the boom of web services,the problem of “information overload”often occurs.For example,in e-commerce platforms,how users should choose from a huge list of items,and what information should be presented to users by companies.These problems have attracted widespread attention among experts and scholars.Recommendation Systems(RS)came into being,which have played a vital role in handling information overload,and have been widely applied in various fields and have achieved huge economic and social benefits.Collaborative Filtering based on Matrix factorization is an effective algorithm for achieving recommendations.It factorizes a useritem preference matrix to learn latent features and realizes personalized recommendation.However,large-scale users and items slow down the online recommendation stage of the recommendation system,causing serious storage and retrieval efficiency problems.Hashing technology projects the features of users and items into the unified hash codes to recommend in Hamming space with high storage efficiency and computational efficiency,which is an effective solution to the recommendation efficiency bottleneck.Hashing-based recommendation methods can find approximate top-K preferred items with sub-linear or constant time complexity and greatly improves the efficiency.But there are also the following limitations:(1)The item cold-start problem is challenging because most hashing-based recommendation methods mainly rely on the user-item interactions and the single auxiliary feature,while ignoring the multi-modal features of items.Intuitively,when the recommender systems have no interaction history with the items or lack a specific auxiliary information,the cold-start problem cannot be handled well.(2)Existing hash recommendation methods learn hash codes based on a batch-based learning model.This mode cannot effectively handle large-scale datasets,it consumes large memory,and has low training stream data efficiency.In response to these two issues,the research content generally contains two parts:(1)To solve the problem of cold start in the recommender system,an effective binary multimodal matrix factorization for fast item cold-start recommendation is proposed.This method proposes an efficient consensus multi-modal mapping to transform the heterogeneous multi-modal features to the unified factors by exploiting the complementarity of multiple modalities.Then,binary matrix factorization is simultaneously performed on the multi-modal features of items and the past user preferences to learn compact binary codes of user/item in a common Hamming space.Besides,inspired by the observation that similar instances often have similar binary codes within a short Hamming distance,this method formulate a semantic structure regularization term to preserve the similarities of items during the binary embedding process.(2)To solve the problem of online learning in the recommender system,an effective online discrete content-aware matrix factorization is proposed.This modal designs a flexible matrix factorization framework for online learning,which can adopt a dynamic update strategy to support real-time online learning when streaming data arrives.And it studies how to hash users and items at the presence of their respective content information for fast recommendation.Besides,to derive compact and informative binary codes,this method adds balanced and de-correlated constraints on binary codes of users and items.This method develops an effective Discrete Coordinate Descent(DCD)approach to tackle the formulated discrete hash optimization problem.Experiments on three publicly available real-world datasets demonstrate the superiority of the proposed approaches against state-of-the-art methods.