Saturday 01 February 2025
The quest for more accurate credit scoring models has led researchers to explore new approaches in machine learning. A recent paper proposes a hybrid model that combines the strengths of two existing methods, KAN and gMLP, to improve performance on large-scale financial tabular data.
Traditional tree-based models like LightGBM are widely used for credit scoring due to their efficiency and effectiveness. However, as data sizes grow, these models may not be able to keep up with increasing complexity. In contrast, deep learning methods have shown promise in handling high-dimensional data, but often struggle with interpretability and require extensive feature engineering.
The proposed hybrid model, TKGMLP, aims to bridge this gap by integrating the advantages of both KAN and gMLP. KAN is a neural network based on the Kolmogorov-Arnold representation theorem, which reduces the dependence on linear weight matrices. gMLP, on the other hand, incorporates gating mechanisms to enhance expressive power by controlling information flow.
The authors designed a novel feature encoding method for numerical features, dubbed Quantile Linear Encoding (QLE), which transforms each feature into a uniform distribution and eliminates extreme values. This approach outperformed traditional methods like centered log ratio and quantile transformation.
Experiments on a credit scoring dataset demonstrated the effectiveness of TKGMLP. The model showed significant improvements over LightGBM as data size increased, with a 1.76% increase in the KS value (a key metric for credit scoring). QLE also outperformed other encoding methods.
The authors’ findings suggest that TKGMLP can be a powerful tool for credit scoring and potentially other applications where large-scale tabular data is involved. By combining the strengths of KAN and gMLP, this hybrid model may offer a more accurate and efficient solution than traditional tree-based models or deep learning methods alone.
The use of QLE as an encoding method is particularly noteworthy, as it can be applied to various numerical features in tabular data. This approach could potentially benefit other applications beyond credit scoring, such as customer churn prediction or recommendation systems.
As the financial industry continues to rely on machine learning for decision-making, innovative approaches like TKGMLP and QLE will play a crucial role in driving progress. By exploring new methods and combining existing techniques, researchers can develop more accurate and efficient models that better serve the needs of businesses and individuals alike.
Cite this article: “A Hybrid Approach to Credit Scoring: Combining Strengths of KAN and gMLP”, The Science Archive, 2025.
Credit Scoring, Machine Learning, Hybrid Model, Kan, Gmlp, Tkgmlp, Qle, Feature Encoding, Numerical Features, Tabular Data
