Dual-target cross-domain recommendation tasks aim to transfer knowledge between the source and target domains to help the mutual recommendation results in both domains. However, learning the user and the item features with a unique feature extraction and aggregation scheme always causes suboptimal results due to their significantly different knowledge transfer patterns. Meanwhile, many existing methods are “black-box”, which cannot provide insightful interpretations for users. To this end, we propose a Meta-path Enhanced Heterogeneous Graph Neural Network (MEH-GNN) for dual-target cross-domain recommendation. First, we conduct a node-specific embedding paradigm to mitigate the knowledge conflicts from different domains, and introduce a new shared Bridge-BiLSTM aggregator to gather information from neighbor nodes. Besides, a hierarchical attention network is performed to improve the bi-directional knowledge transfer by adjusting the weights of instance-level, path-level and domain-level attentions adaptively to learn enhanced representations of the users and items. MEH-GNN considers both domain-specific and domain-shared features when fusing features and relies on the hierarchical attention mechanism to balance between them. Finally, our intra- and inter-domain meta-paths sampled in the cross-domain heterogeneous graph provide path-based interpretability. Extensive experimental results demonstrate that MEH-GNN outperforms existing state-of-the-art methods in recommendation accuracy and can achieve path-based interpretability.