In recent years, lots of techniques applied and optimized to make the recommender system better. Most of them mainly focus on the target interaction between users and items or a part of auxiliary information but hardly leverage other useful information relevant to customer transactions. In this paper, we want to propose a new model named Heterogeneous Neural Collaborative Filtering (HNCF) for learning recommender systems from two important parts: multi-aspects and multi-behaviors. The HNCF algorithm proposed is divided into four parts: Commuting similarity matrix, Multi-Layer Perceptron, Fusion by the Attention Mechanism, Multi-behavior Prediction. The model exploits characteristics of customers and properties of products from different aspects besides the aspect of purchase by building meta paths then commuting similarity matrices. Aspect-level latent factors fusion gives results as factors representing each user and item. These factors synthesizing with multi-behaviors prediction layers is the highlight of the model to make a better recommender system.