Traditional recommender systems that are based on CF (Collaborative Filtering) usually suffer from the user-item interaction data sparsity problem. With the development of end-to-end models founded on GNNs, the sparsity issue can be addressed by introducing additional sources of information such as knowledge information. However, these models are insufficient to fuse multi-source information. To fill this gap, in this paper, we propose an end-to-end GNN based model called KCIRec, which fuses both K _ nowledge graph C _ ollaborative information and user-item I _ nteraction information for Rec _ ommender system.Technically, a type of two-channels’ information propagation and aggregation mechanism are conceived to generate the representation of user-item interactions graph and collaborative knowledge graph respectively.In addition, we design an attention mechanism to adaptively fuse the collaborative information and knowledge graph information extracted from the above two graph. Extensive experiments on three real-world datasets show improvements of our proposed KCIRec model over thestate-of-the-art methods such as KGNN-LS, KGAT, and CKE. The promising results show that the proposed KCIRec is able to effectively fuse knowledge graph information and improve recommender systems’ accuracy.