<p>Kerogen type classification is central to hydrocarbon exploration because it constrains the petroleum generation potential of source rocks. However, frontier lacustrine basins, kerogen typing is often hindered by overlapping Rock-Eval signatures, class imbalance, and subjective threshold interpretation. This study introduces an explainable ensemble machine-learning approach for kerogen classification in the Nkhata Basin, Malawi. Using 153 Rock Eval and Total Organic Carbon (TOC) samples, predictive accuracy is integrated with geochemical interpretability under data-limited, thermally immature basin conditions. Following median imputation, Box–Cox normalization, and domain-driven feature engineering, eleven supervised algorithms were evaluated using stratified five-fold cross-validation. Gradient Boosting achieved the most stable performance (macro-F1 = 0.9349 ± 0.0399; MCC = 0.9569), while Random Forest attained perfect test-set metrics but showed higher cross-validation variance (± 0.138), indicating reduced fold stability. Imbalance-aware evaluation using macro-averaged metrics and confusion matrix analysis confirmed reliable classification of minority kerogen classes. Residual misclassification occurred predominantly between Type II and Type II/III kerogen (68% of errors), reflecting genuine geochemical transition rather than model instability. Explainable AI analyses using SHAP and LIME demonstrate that Hydrogen Index (HI) and petroleum potential proxies (S2, S1 + S2, S2/S3) account for over 70% of predictive influence, with consistent feature rankings across folds. The convergence of global and local explanations confirms that model decisions are grounded in established geochemical principles rather than spurious statistical correlations. By coupling ensemble learning with interpretable AI, a transparent and interpretable kerogen typing framework applicable in early-stage exploration is developed in frontier basins.</p>