Hong, Chen-HsiangChen-HsiangHongRUEY-SHAN GUOCHIA-LIN CHEN2025-12-102025-12-102025-10https://scholars.lib.ntu.edu.tw/handle/123456789/734501The semiconductor industry, a foundation of the modern economy, demands integrated evaluation approaches that consider financial, technological, and operational dimensions together. However, existing methods often address these aspects in isolation, limiting comprehensive understanding. This study proposes a novel SHAP-value-based performance evaluation framework that combines eXtreme Gradient Boosting (XGBoost) with dynamic network Data Envelopment Analysis (DEA). Unlike traditional uses of SHAP for post hoc interpretation, we employ SHAP as a pre-analysis tool to systematically select variables and allocate stage-specific linkage ratios, objectively quantifying the influence of each dimension on firm performance. Furthermore, by dynamically recalculating SHAP values across different periods, our model captures temporal shifts in factor importance — a capability not addressed in conventional network or dynamic DEA approaches. This dynamic design makes the framework particularly well-suited for analyzing changes across different market conditions, especially during downturns, where successful firms may adopt distinct strategic decisions compared to failing ones. By tracking how the importance of financial, technological, and operational variables evolves over time, the model provides insights into critical success factors under varying external shocks. While this study uses the Covid-19 pandemic as a demonstration case, the framework may be applicable to other disruption scenarios. Overall, this research bridges critical methodological gaps and offers a robust tool for strategic decision-making in highly volatile industries.Data envelopment analysisMachine learningExtreme gradient boostingSHAPPerformance evaluation[SDGs]SDG9journal article10.1016/j.cie.2025.111284