Mutually Promoted or Mutually Restricted? A Multi-level Co-adaptation Regularized Deep Learning Method for Multimodal Data Based Default Risk Prediction.