<p>Heterogeneous graphs can represent complex relationships in the real world through multiple types of nodes and connecting edges. The goal of heterogeneous graph representation learning is to embed the structural and semantic information of the target nodes into a low-dimensional vector representation, which can facilitate the analysis of downstream tasks (i.e., node classification, and node clustering). However, the majority of existing methods for learning heterogeneous graph representations focus primarily on single-view information, overlooking the multiview properties inherent in such graphs. To address these limitations, we propose multiview fusion neural networks (MVF-NNs) for heterogeneous graph representation learning. This model can be divided into two parts: a multiview feature extraction module and a multiview fusion module. The multiview feature extraction module learns representations from three views: a feature view, a metapath view, and a structure view. The multiview fusion module fuses representations of these views by using an attention mechanism, multihead self-attention, with an MLP and pairwise combinations to achieve the final weighted aggregation. To validate the practical effectiveness of the proposed method, we conduct node classification and clustering experiments on four real-world datasets. The results demonstrate average improvements of approximately 2% in both tasks compared to state-of-the-art models. The practical benefits of these improvements translated into real-world applications are manifested in the form of improved effectiveness of recommender systems and improved accuracy of social network analysis.</p>