From cars to clinics: interpretable machine learning for luxury automotive lifecycle and clinical case studies

This thesis investigates how interpretable machine learning can drive strategic success in the luxury automotive sector and beyond by analyzing the customer lifecycle through three interconnected dimensions: personalization, retention, and residual value. Developed within a leading manufacturer’s ecosystem, the research first introduces a hybrid recommender system for highly configurable vehicles that achieves high precision and diversity through a scalable XGBoost architecture. To address customer churn during long waiting periods, a second contribution utilizes a CatBoost classifier enhanced by SHAP explanations and Large Language Models to provide actionable insights for dealerships. The third automotive pillar focuses on residual value estimation by integrating vehicle attributes with macroeconomic indicators, achieving a low mean absolute percentage error while introducing a standardized vehicle concept to isolate genuine market trends. Beyond the automotive domain, the methodology extends into high-stakes clinical and sports contexts characterized by data scarcity and heterogeneity. In neurology, the framework predicts GBA1 mutation status in Parkinson’s disease patients with 73% accuracy, identifying clinical patterns that support targeted genetic testing. In sports science, the model analyzes footballers’ biomechanics to identify movement phenotypes, achieving a macro-F1 \approx 0.92$ and providing actionable insights for preventing ACL injuries. Ultimately, this work demonstrates that encoding domain structure and prioritizing transparency allows deployment-oriented machine learning to function effectively even when data is scarce, effectively closing the loop between complex data insights and operational outcomes.