Recommendation system is a passive user application designed to mitigate information overload. By reducing the time users spend filtering content and uncovering potential users for enterprises, it has emerged as a technology garnering considerable attention in both research and industry circles. The recent strides made in Graph Neural Networks (GNNs) have opened doors to notable enhancements in recommendation system. Nevertheless, current GNNs-based recommendation systems face several challenges, such as insufficient node representation learning, underutilization of rich interactive data, and ineffective handling of sparse datasets. In view of the above problems, this paper introduces a novel model named MB-DCGBN(Multi-Behavior Recommendation with Dual Channel Behavior Perception Based on Graph Attention Network). It utilizes graph attention network to capture complex higher-order relationships between different nodes. In view of the diversity of interactive behaviors in real application scenarios of recommendation system, MB-DCGBN implements a hierarchical graph attention network module and an explicit edge embedding-integrated graph attention network module. These components aim to capture the high-level impact of different behaviors on users and items, and enhance the semantic expression ability. By addressing the challenges encountered by conventional methods, MB-DCGBN demonstrates significant performance improvements over baseline models.