An Empirical Evaluation of Explainability Drift Under Post-Training Quantization of Machine Learning Models

Post-training quantization (PTQ) is widely adopted to reduce the computational and memory footprint of machine learning models deployed in resource-constrained environments. While its impact on predictive performance has been extensively studied, its implications for model explainability remain insufficiently understood. Explanations are often assumed to remain reliable as long as predictive accuracy is preserved. This assumption, however, has rarely been examined empirically. In this work, we present a systematic empirical study of explainability stability under PTQ across three representative model families, namely, Logistic Regression, Random Forest, and XGBoost, using the IEEE-CIS Fraud Detection dataset. Explanation stability is evaluated using complementary global and local attribution metrics, including feature rank correlation, mean attribution drift, and instance-level cosine similarity of SHapley Additive exPlanations explanations. Models are trained once in full precision and subsequently quantized to INT8 and INT4 without retraining to isolate numerical precision effects. The results indicate that predictive performance and explanation stability are not inherently aligned. While some models retain stable predictive metrics under reduced precision, measurable attribution drift may still occur. In particular, aggressive quantization can induce substantial instability in explanation behavior even when aggregate performance remains unchanged. These findings present that explainability cannot be inferred solely from accuracy metrics and should be explicitly evaluated when deploying quantized models in high-stakes decision-making systems.