Credit risk assessment using the factorization machine model with feature interactions

Credit risk, the probability of borrower default, is crucial for financial institutions. Historical failures highlight its significance, with events like the collapse of Bank Herstatt in 1974 and the impact of the 2008 financial crisis emphasizing the need for robust credit risk assessment. Traditional methods such as logistic regression (LR) and linear discriminant analysis (LDA) rely on linear assumptions, limiting their ability to capture complex non-linear relationships present in credit risk data. While machine learning techniques like Support Vector Machines (SVM), k-Nearest Neighbors (kNN), and Artificial Neural Networks (ANN) offer improvements, they may not fully capture feature interactions. The factorization machine (FM) model, proposed by Rendle in 2010, addresses this gap by explicitly modeling feature interactions. This study investigates whether the FM model can improve credit risk assessment by effectively capturing these interactions, using real-world datasets to compare its performance with traditional models. The research aims to provide empirical evidence of FM's effectiveness and explore potential avenues for enhancing its performance in credit risk assessment.

Credit risk assessment has evolved from parametric statistical models like LR and LDA to machine learning techniques such as SVM, kNN, and ANN. LR and LDA, while capable of estimating default probabilities, assume feature independence, limiting their accuracy. Non-parametric approaches like the historical random forest (HRF) have shown promise in outperforming parametric methods for SMEs. SVM, with its ability to handle high-dimensional data and non-linear relationships, has achieved better performance compared to other classification methods in credit risk management. However, these traditional models often struggle to effectively capture the complex interactions between different features that significantly influence credit risk. The factorization machine (FM) model is presented as a superior alternative that addresses this limitation by explicitly modeling these interactions.

This research uses four real-world credit datasets from the UCI Machine Learning Repository: Bank Marketing, Credit Approval, German Credit, and Statlog (Australian Credit Approval). Data preprocessing involved handling missing values (through deletion or imputation) and outliers (through verification and replacement or removal). One-hot encoding transformed categorical features into numerical values. The study compared the performance of the FM model with LR, SVM, kNN, and ANN. The datasets were split into training (75%) and testing (25%) sets. Model parameters were tuned using cross-validation. Model performance was evaluated using several metrics: accuracy (ACC), Matthews Correlation Coefficient (MCC), precision (PRE), recall (REC), F-score, true positive rate (TPR), true negative rate (TNR), false negative rate (FNR), false positive rate (FPR), area under the ROC curve (AUC), and G-mean. These metrics provide a comprehensive assessment of the models' performance across different aspects of credit risk prediction, including the ability to correctly identify both defaults and non-defaults.

Across all four datasets, the FM model consistently outperformed LR, SVM, kNN, and ANN in terms of several evaluation metrics. The FM model showed the highest ACC, MCC, AUC, and G-mean values in most cases. While other models (such as LR and SVM) occasionally excelled in specific metrics (like PRE and TNR), the overall superiority of the FM model was evident. For example, on the Bank Marketing dataset, FM achieved an ACC of 0.9021, MCC of 0.4922, and AUC of 0.7343, surpassing other models. Similar superior performances were observed across the Credit Approval, German Credit, and Statlog (Australian Credit Approval) datasets. These findings consistently demonstrated the FM model's ability to effectively capture feature interactions, leading to improved predictive accuracy and a more comprehensive assessment of credit risk.

The results demonstrate the effectiveness of the FM model in credit risk assessment, highlighting its ability to handle feature interactions and achieve superior performance compared to traditional machine learning methods. The consistent outperformance across diverse datasets suggests that FM's capability to model complex relationships between features is a key factor contributing to its success. This advantage is particularly relevant in credit risk scenarios where numerous factors often interact in intricate ways. While some datasets show exceptional performance by other models in specific metrics, this does not detract from FM's overall superior performance. The findings suggest that the FM model can serve as a valuable tool for financial institutions seeking to enhance the accuracy and efficiency of their credit risk assessment processes.

This study shows that the FM model outperforms traditional machine learning models in credit risk assessment due to its superior ability to capture feature interactions. The consistent superior performance across four real-world datasets underscores its potential for practical application in financial institutions. Future research could explore adapting FM models for online credit risk assessment and optimizing the loss function to further enhance performance.

While the FM model demonstrates superior performance, certain limitations exist. The optimal parameter selection for the FM model may vary across datasets, requiring further investigation for consistent optimal performance across diverse datasets and scenarios. The impact of data imbalance on model performance also requires more detailed analysis. Further research could focus on exploring how FM performs with even larger and more diverse datasets.