Research on Interpretable Loan Approval Identification Using Multi-Dimensional Features

In recent years, as demand for social financing has continued to grow, loans have become an indispensable financial tool for both individuals and businesses. For financial institutions, the quality of loan approval decisions directly affects asset security and operational stability. Against this backdrop, credit scoring models serve as the foundation of automated approval systems by quantifying borrowers' credit risk and providing an objective basis for approval decisions [1]. A robust and accurate model not only improves approval efficiency and reduces default losses but also plays a key role in advancing the intelligent transformation of financial services.

However, traditional loan approval processes have long relied on manual experience, which comes with limitations such as low efficiency and inconsistent standards. While machine learning techniques have been widely adopted to build automated approval models, existing research still has gaps: on one hand, most models focus narrowly on default prediction and fail to adequately support comprehensive approval decisions involving multi-dimensional rules; on the other, many high-performance models are "black boxes" with opaque decision-making logic [2], making it difficult to meet the strict regulatory requirements for fairness and interpretability, and leaving applicants unable to understand the rationale behind approval outcomes. Therefore, exploring loan approval models that combine high predictive performance with strong interpretability has become an important focus for both academia and industry.

In response to these challenges, extensive research has been conducted in academia on loan approval modeling, forming three main directions.

In terms of constructing ensemble learning and complex models, Kokate and Chetty employed a combination of machine learning methods such as gradient boosting, random forests, and decision trees to build a credit scoring model for automated approval, using feature selection techniques to improve model efficiency. Their model demonstrated superior risk discrimination ability and stability on real-world banking data, validating the effectiveness of combining ensemble learning with feature optimization [3]. Further, Uddin et al. proposed a hybrid approach integrating deep learning with ExtraTrees and adopted an ensemble voting mechanism to combine the three best-performing base models, achieving an accuracy of 87.26% in bank loan default prediction, highlighting the significant improvement of ensemble strategies on classification performance [4]. In line with this, Perera and Premaratne developed a stacking ensemble model based on a voting mechanism, further confirming the practicality and stability of ensemble learning in credit risk assessment [5]. Lakshmi and Rao constructed a loan default prediction model by integrating algorithms such as naive Bayes, decision trees, and multi-layer perceptrons, training on Kaggle historical data and ultimately achieving 90% accuracy, further demonstrating that ensemble learning methods can effectively support banks in making loan approval decisions [6].

In terms of benchmarking and optimization of traditional machine learning models, a substantial body of research has focused on providing benchmarks for model selection through systematic comparisons. Singh employed logistic regression, random forests, and support vector machines to model loan approval, achieving an accuracy of 78.785% on historical data [7]. Nureni and Adekola systematically compared eight machine learning algorithms and found that logistic regression performed particularly well in terms of accuracy (up to 83.24%) and sensitivity (up to 97.76%) [8]. Tumuluru et al. pointed out that among models for default prediction based on historical customer data, the random forest algorithm achieved the best accuracy [9]. Viswanatha conducted a comprehensive comparison of algorithms including random forests, naive Bayes, decision trees, and KNN, with naive Bayes achieving the highest accuracy at 83.73% [10]. A systematic comparison by Sarkar et al. also indicated that random forest and logistic regression models slightly outperformed other compared algorithms in terms of overall performance. These extensive comparative studies provide empirical support for model selection across different scenarios [11].