Research On Side Information Augmented Collaborative Recommender Systems

Recommender systems have been widely used in e-commerce,social media,education and scientific research,etc.,it can deal with the problem of information overload effectively.However,the problem of data sparsity has always restricted the effect of the recommender systems.Using side information(content information,user social network)can effectively alleviate the side effects of data sparseness,but how to effectively use this information to enhance the performance of recommender systems is still an open issue.In addition,with the advancement of technologies such as deep learning and metric learning,research on new recommender systems via new technologies and side information has become a new challenge.Seeking to the two main tasks of rating prediction and top-n recommendation in recommender systems,this thesis respectively exploits reviews and social network to construct side information augmented recommender systems.To solve the problem of rating prediction,a dual-regularized matrix factorization(DRMF)with deep neural network is proposed.DRMF firstly uses the deep neural network which consists of convolution neural network and gated recurrent unit to process the aggregated reviews of users and items respectively,and obtains two independent distributed representations to represent users and items.Then,these representations serve to constrain the embedding of users and items in the process of probabilistic matrix factorization.The joint parameter optimization of DRMF model is realized by alternate training,that is,each time the parameters of matrix factorization are fixed,the parameters of deep neural network components are optimized by back propagation,and then the latter parameters are fixed to optimize the parameters of matrix factorization.The intensive experiments have proved that DRMF can significantly reduce the error of rating predictions.As for the task of top-n recommendation,this thesis proposes a social-regularized collaborative metric learning(SRCML)method.SRCML defines a similarity function to measure user similarity by integrating social preference and rating preference.Then,it constructs regularization component to reorganize the loss function of collaborative metric learning,so that users with similar interests come closer to each other in metric space,and the metric function to measure the distance between users and items is optimized.Finally,a model learning method based on gradient descent is presented.A large number of experiments prove that SRCML can dramatically improve the accuracy and recall rate of recommendations.