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

Authors

Gang Chen, School of Management, Fudan UniversityFollow
Tian Lu, Carnegie Mellon UniversityFollow
Chenghong Zhang, Fudan UniversityFollow

Abstract

The proliferation of social media along with AI techniques have flourished the use of multimodal data to default risk prediction. In this study, we submit a pair of opposite effects for multimodal data based default risk prediction, namely, mutual promotion effects and mutual restriction effetcs. Given such a premise, we firstly extract hard features, social media textual features, and facial attractiveness features for each of the borrowers, and then we propose a novel multi-level co-adaptation regularized multimodal deep learning method. Our proposed method is capable of identifying mutually promoted patterns among multimodal features, besides, our method opens up a novel avenue for balancing the trade-off between feature-wise co-adaptation and modality-wise co-adaptation. Moreover, we theoretically prove that by virtue of our method, multimodal features with multi-level co-adaptations can be automatically detected. The experimental resuts show that our method beats existing methods by a significant margin.

Recommended Citation

Chen, Gang; Lu, Tian; and Zhang, Chenghong, "Mutually Promoted or Mutually Restricted? A Multi-level Co-adaptation Regularized Deep Learning Method for Multimodal Data Based Default Risk Prediction" (2020). PACIS 2020 Proceedings. 74.
https://aisel.aisnet.org/pacis2020/74