A Hybrid Adaptive Leadership GWO Optimization with Category Gradient Boosting on Decision Trees Algorithm for Credit Risk Control Classification

Suihai, Chen and Chih How, Bong and Po Chan, Chiu (2024) A Hybrid Adaptive Leadership GWO Optimization with Category Gradient Boosting on Decision Trees Algorithm for Credit Risk Control Classification. PhD thesis, Universiti Malaysia Sarawak.

PDF THESIS PHD_Chen Suihai.pdf Restricted to Registered users only Download (5MB) | Request a copy

Abstract

With the rapid development of financial economy in our country, credit business has become the main business of banks and financial companies. The quality of customer credit directly affects the business performance of financial companies. The traditional risk control model cannot identify good customers and bad customers well, so a new high precision risk control algorithm model is needed to solve this problem. CatBoost algorithm has high accuracy and takes into account the efficiency of the model, which can effectively solve such problems. However, compared with the traditional risk control algorithm (logistic regression algorithm), CatBoost algorithm also needs to have the advantages of high efficiency, low algorithm complexity and strong interpretable ability. Therefore, this study first reduces the dimension of the variable features (deleting the variable features with low correlation), and then uses two different datasets for experimental comparison and verification, which proves that the dimension reduction can improve the efficiency of the algorithm. Secondly, an improved CatBoost algorithm (EBGWO-CatBoost) was proposed, which was a combination of improved GWO algorithm (EBGWO) and CatBoost algorithm, and the optimized GWO algorithm was used to offset the defects of CatBoost algorithm in parameter tuning. The combined EBGWO-Catboost algorithm model is structured into four distinct steps. Initially, the Grey Wolf algorithm (GWO) is leveraged to exploit its advantages in identifying optimal parameters. Subsequently, the optimal parameters of the CatBoost algorithm are determined to streamline computational resource usage and reduce model complexity. However, optimizing the balance between exploration and exploitation within the GWO algorithm itself is imperative. Hence, the second step entails enhancing the GWO algorithm, referred to as the EBGWO algorithm. This enhanced version of the Grey Wolf Optimization algorithm possesses robust global search capabilities and helps alleviate some of the local convergence issues inherent in the original GWO algorithm. It can effectively enhance the predictive accuracy and execution speed of the CatBoost algorithm model. The third step involves applying the new algorithm to the risk control model for testing and comparison, resulting in the conclusion that the model established by the EBGWO-Catboost algorithm exhibits more advantages compared to models built by other algorithms. Thirdly, this study uses SHAP framework to improve the interpretability of the new algorithm (EBGWO-CatBoost), and solves the problem of the weak interpretability of the new algorithm. Finally, a summary of the entire document is provided, along with suggestions for future work and extensions.

Actions (For repository members only: login required)

View Item