Modeling Relationships Between Users And Items With Nonlinear Ways

With the growing volume of online information,recommender systems have been an effective strategy to overcome information overload.In addition,modeling relationships between users and items plays a key role in recommender systems.Unfortunately,most of existing method models it by a linear way,which may limit the methods' performance.Although some recent work has employed deep learning for capturing the nonlinear relationships between users and items.However,two potential problems may emerge when the deep neural work is exploited.The fundamental problem is that the complexity of the algorithm grows significantly with the increment in the depth of the neural network.The second one is that a deeper neural network may undermine the accuracy of the algorithm.For another,the multi-criteria ratings are utilized in improving the accuracy of predicting overall ratings.One of the most important challenges of multi-criteria recommender systems is how to exploit latent relationship between multi-criteria ratings and overall ratings.Although the transfer learning technique is successfully applied in social recommender systems,Point-of-Interest recommender systems and content-based recommender systems,the study of transfer learning in the multi-criteria recommender systems still attracts little attention.In order to address those limitation of recommender systems.In this paper,we respectively exploit the heterogeneous neural network and auxiliary information on modeling the relationship between users and items.Specifically,main contribution of our work is summarized as follows:1.We propose a hybrid neural network that consists of the global neural network and the local neural block to learn the nonlinear relationships between users and items.Three architectures are used to build the local neural block.Moreover,we combine the hybrid neural network with the GMF model.Our method can adopt more neural layers,which makes the information learnt by our method is more accurate.The experimental results on real datasets reveal that our method is superior to the state-of-the-art methods in terms of the item ranking.2.We propose a collective factor model to integrate contributions of overall ratings and multi-criteria ratings.Specifically,overall ratings are considered as the target data and multi-criteria ratings are regarded as the auxiliary data.We combine loss functions of overall ratings and multi-criteria ratings.Parameters of the model are learned by employing both overall ratings and multi-criteria ratings.It is verified by the experimental result on several benchmark datasets that the proposed method is superior to baseline methods in terms of overall rating predictions.