Data-Driven Insights Into Underdeveloped Regencies: SHAP-Based Explainable Artificial Intelligence Approach

Abstract

Classification analysis in high-dimensional data presents significant challenges, particularly due to the presence of complex non-linear patterns that traditional methods, such as logistic regression, fail to capture effectively. This limitation is often reflected in relatively low model accuracy. One approach to addressing this issue is through machine learning-based classification methods, such as Random Forest and Support Vector Machine (SVM). While these models generally achieve higher accuracy than logistic regression, their black-box nature limits interpretability, making it difficult to explain their classification decisions. As machine learning models continue to advance, interpretability has become a crucial concern, especially in data-driven decision-making. Post-hoc explainable artificial intelligence (XAI) techniques offer a viable solution to enhance model transparency. This study applies SHAP to machine learning models to gain insights into the underdevelopment status of regencies in Indonesia. The results indicate that SVM outperforms both logistic regression and Random Forest. SHAP values estimated from SVM, using various permuted variable subsets, exhibit stability. Clustering analysis identifies five optimal clusters of underdeveloped regencies. Based on average SHAP values, underdevelopment alleviation strategies should focus on social factors (Cluster 1), infrastructure (Cluster 2), accessibility (Cluster 3), and a combination of infrastructure, accessibility, education, and healthcare (Cluster 4), while Cluster 5 requires improvements in accessibility and economic conditions.