<p>The prediction of credit risk for small and medium-sized enterprises (SMEs) is a critical topic that has attracted significant attention. Recently, complementary multi-source data is explored to improve the performance of prediction models. However, traditional methods tend to concatenate the multi-source data into a single input, which may destroy the inherent structure of heterogeneous data and cause curse of dimensionality. In this study, we propose a novel method called multi-view learning with hierarchical attention mechanism (MVL-HA) to effectively integrate multi-source data for the credit risk prediction of SMEs. Specifically, distinct neural networks are employed based on the characteristics of each data source, while a hierarchical attention mechanism is introduced to deal with features within each view and across different views at a fine-grained level. Extensive experiments on a Chinese dataset demonstrate that the proposed method achieves excellent performance compared to the baseline models, particularly in dealing with imbalance problem in credit risk prediction.</p>