With the ever growing number of choices in many online services,the amount of data on the Internet presents an explosive growth trend,which makes it difficult for users to effectively obtain information of interest.Recommender system is an impor-tant tool to help users find items that meet their interests and relieve the information overload problem.Matrix factorization is currently one of the frontiers in the domain of recommender systems.The aim of this algorithm is to find two latent factor matri-ces in the same latent space,which has many advantages such as good theoretical basis and high prediction accuracy.However,matrix factorization still has some problems such as time-consuming process to choose an optimal learning rate,poor parallelism on skewed rating data and unsatisfactory results on sparse rating matrix.This paper focuses on matrix factorization,analyzes the existing problems and puts forward the corresponding improved algorithms.The main work and contributions are as follows.We propose an adaptive learning rate schedule for stochastic gradient descent to matrix factorization.Stochastic gradient descent is an effective algorithm to solve ma-trix factorization problem.However,the performance of stochastic gradient descent de-pends critically on how learning rates are tuned throughout the training process.This pa-per presents a novel per-dimension learning rate schedule called AALRSMF(An Adap-tive Learning Rate Schedule for Matrix Factorization).This schedule relies on local gradients,requires no manual tunning of a global learning rate,and shows to be robust to the selection of hyper-parameters.We conduct extensive experiments and the results demonstrate that the proposed schedule achieves a faster convergence speed and shows insensitive to the hyper-parameters selection compared with other schedules on matrix factorization.We propose a fast and robust parallel SGD matrix factorization algorithm.Re-search on parallelization for matrix factorization has always been a hotspot,efficient parallel SGD matrix factorization algorithms have been developed for large matrices to speed up the convergence of factorization.However,when the rating matrix is skewed,their performances are unreliable.This paper presents a novel parallel SGD matrix fac-torization algorithm called KDMF(KD-tree for Matrix Factorization),which is robust to skewed matrices and run efficiently on shared memory systems.KDMF uses kd-tree for partitioning the rating matrix and minimizes the cost for scheduling parallel SGD updates on the partitioned regions by exploiting partial match queries.Thereby,KDMF produces reliable results efficiently even on skewed matrices.From our extensive eval-uations,KDMF significantly outperforms several state-of-the-art parallel algorithms.We propose a novel document context-aware matrix factorization algorithm.Ma-trix factorization(MF)is an effective approach commonly used by many recommender systems.Conventional MF-based methods use the ratings given to items by users as the sole source of information to make recommendations.However,the ratings are often very sparse in many online applications,causing MF-based methods to degrade signif-icantly in their recommendation performance.To handle the sparsity problem,several past works on the problem of document context-aware recommendation have been pro-posed to improve the rating prediction accuracy by additionally utilizing item content information such as tags,reviews,and descriptions,etc.Nevertheless,due to the inher-ent limitation of word-based methods,they have difficulties in effectively utilizing con-textual information of item content documents,which leads to shallow understanding of the documents.This paper presents a novel document context-aware method called CharConvMF(Character-level Convolutional Matrix Factorization),which jointly per-forms deep convolutional neural network representation for the item content informa-tion and matrix factorization for the rating data.Consequently,CharConvMF treats item content documents as raw signals at character-level to capture contextual features of documents and incorporate contextual features into matrix factorization framework to further enhances the rating prediction accuracy.Our extensive evaluations on three real-world datasets show that our proposed method leads to a significant improvement over several state-of-the-art recommendation methods even when the rating data are extremely sparse. |