<p>Knowledge graph (KG) is a highly structural knowledge system, which is widely used for reinforcing the performance of downstream applications. Recently, to enhance recommendation performance, knowledge-driven recommender system has emerged as a research hot-spot in KG applications. However, the existing models are lack of the mutual and deep features mining from it, thus it has underused. For this research, two components are formulated to manage it. They are mutual information captor (MIC) and knowledge aware learner (KAL) in our model, respectively. In MIC, an attention mechanism is used for mutual information between entities in KG and items in recommander system capturing to obtain more features in the form of multi-task feature learning methods; In KAL, a neural network model based on deep factorization is proposed to further mine the deep features between users and items, facilitating knowledge extension for recommendations. The extensive experiments demonstrate that the proposed model has achieved promising performance over almost all the state-of-the-art existing models in this research line.</p>