在金融機構授信決策中，信用風險管理扮演關鍵角色，能否準確辨識高風險借款人並建構可靠的信用評分模型，直接影響授信品質與風險控管成效。然而，傳統機器學習演算法在處理信用資料時常面臨兩大挑戰，一是高維度資料中的冗餘與雜訊特徵易干擾預測效能，二是違約樣本比例偏低所導致的類別不平衡問題，易使模型偏向多數類別而忽略少數關鍵樣本。此外，現有研究多傾向採用單一特徵選取方法，未能充分整合不同方法間的互補資訊。 為解決上述問題，本研究結合三種特徵選取方法與兩種融合策略，採用五個公開信用風險資料集，並對五種分類器進行系統性實驗。透過比較分析，本研究評估單一特徵選取與特徵層級融合之預測表現差異，並進一步引入決策層級融合機制，探討整合多個分類器對模型預測能力之提升幅度。 研究結果顯示，整合多種特徵選取方法所篩選之特徵子集，相較於單一特徵選取，能有效提升模型之預測效能與少數類別辨識能力。本研究提供一套系統性的特徵選取與融合策略實驗框架，期望協助金融機構更精確地辨識高風險借款人，降低信用損失。;Credit risk management is critical to lending decisions in financial institutions， where accurately identifying high-risk borrowers and constructing reliable credit scoring models directly affects lending quality and risk management effectiveness. Traditional machine learning algorithms face challenges with credit data， including redundant and noisy features in high-dimensional data that degrade predictive performance， and class imbalance from low default rates that biases models toward the majority class while neglecting the critical minority class. Most existing studies rely on a single feature selection method， leaving complementary information across different methods insufficiently integrated.
To address these issues， this study combines three feature selection methods with two fusion strategies， using five publicly available credit risk datasets and conducting systematic experiments on five classifiers. Through comparative analysis， this study evaluates the performance differences between individual feature selection and feature-level fusion， and further introduces a decision-level fusion mechanism to investigate the benefits of integrating multiple classifiers on model predictive capability.
The results show that integrating the feature subsets selected by multiple feature selection methods effectively improves the model′s predictive performance and minority-class identification compared with a single feature selection method. This study provides a systematic experimental framework of feature selection and fusion strategies， with the aim of helping financial institutions identify high-risk borrowers more accurately and reduce credit losses.
