Heterogeneous Representation Decomposition-Fusion Network with multi-resolution wavelet transform for credit scoring

The accuracy of credit scoring directly influences credit decision-making and the profitability of financial institutions. Traditional credit scoring models typically employ a straightforward approach by directly concatenating continuous and discrete features. However, these methods fail to account for the complex interactions between features, particularly the multi-scale relationships inherent in the interplay between continuous and discrete data. As a result, these models struggle to capture the full spectrum of borrower credit behavior, limiting their performance and robustness. To overcome this limitation, this paper proposes a novel Heterogeneous Representation Decomposition-Fusion Network (HRDN) that incorporates multi-resolution wavelet transform for credit scoring. Specifically, our model first separately extracts features from both discrete and continuous tabular data, then uses a feature pyramid alignment mechanism to fuse these features for credit scoring estimation. To explore the complex nature of continuous data, we design a multi-scale credit representation decomposition (MCRD) module with wavelet transform to decompose the continuous data into multi-resolution feature representations, leading to a more refined characterization of inherent structure and dynamic properties in financial data. Moreover, the feature pyramid alignment module is devised to fuse multi-scale representations for model’s capacity enhancement and robustness improvement. Experimental evaluations on three publicly available datasets demonstrate the effectiveness of the HRDN model, achieving AUC improvements of 1.68%, 0.83%, and 1.17% over state-of-the-art methods. Our model offers strong technical support for credit risk assessment and decision-making in financial institutions, with promising potential for widespread application in the financial industry.