Recommender systems have become a pervasive and effectual means to cope with the boosted complexity of decision-making process resulting from information overload. The key factor confronting recommendation tasks is to accurately characterize user preferences which are not only intrinsically dynamic and evolving, but also associated with multiple features and contexts. Sequential models especially Recurrent Neural Networks (RNNs) have become a powerful tool to address the issue. However, most of current RNN methods select the sequence of items that the user interacted with in the past to represent the current user state without considering contextual information associated with users and items simultaneously. Besides, the existing methods are hard to access and store the importance information of user representation. In this paper, we put forward two types of Gated Recurrent Units (GRU) models, i.e., multi-feature linear fusion GRU and multi-feature user integration GRU, to address the above drawbacks and thus help recommend more personalized next items to users. The proposed models employ multi-feature fusion technology including feature concatenation and one-dimensional Convolutional Neural Network (1D CNN) to generate input of GRU for learning more feature interaction information. Experiments on two real-world datasets demonstrate the performance gains of the proposed model compared with state-of-the-art. Further ablation studies verify the impact of different features for two models.