<p>The discovery of novel, high-performance thermoelectric materials is a critical challenge in materials engineering. To accelerate this process, we establish an interpretable machine learning workflow that provides data-driven guidance for materials design. Our approach begins with constructing a comprehensive descriptor set from fundamental atomic features, followed by a rigorous, three-stage feature selection process to identify the most critical performance-determining properties. Using these selected features, we train and evaluate several regression models, identifying LightGBM as the optimal model with a high coefficient of determination (R²= 0.90) on a held-out test set. The model’s robustness and predictive accuracy are further confirmed through rigorous leave-one-out cross-validation. Crucially, we employ SHAP (Shapley Additive Explanations) to interpret the model, revealing the complex, nonlinear relationships between material features and thermoelectric performance. This workflow not only accurately predicts the performance of known materials, but also provides actionable insights for designing new compositions with enhanced properties, thereby offering a validated and practical tool to guide experimental materials discovery.</p>