Slow cook-off （SCO） response assessment and formulation optimization of melt-cast explosives are challenged by the high cost of experiments and the resulting limited sample size. To address this problem， a data-driven model linking formulation composition to SCO response level was developed. Twenty-two formulations were prepared using a fixed casting process and subsequently evaluated by SCO tests under fixed conditions. The mass fractions of 12 components were used as input features， and the SCO response level was used as the class label. To alleviate minority-class scarcity and class imbalance， SMOTE-based oversampling was applied to construct a class-balanced dataset containing 32 samples for model training and evaluation. Using stratified sampling and three-fold cross-validation， ordinal logistic regression， multinomial logistic regression， random forest， and support vector machine were comparatively evaluated. Component contributions were then analyzed to improve model interpretability. Results show that non-ordinal models outperform ordinal models， and the random forest model achieves the best performance （accuracy 0.79， precision 0.78， and F1-score 0.75）. Misclassifications occur mainly between deflagration and explosion. The feature-importance ranking is broadly consistent with engineering experience， with component IDs C， I， and K identified as dominant， followed by G， A， and H. Based on these findings， a forward design framework termed “profile scanning-small-step validation-model updating” is proposed. By visualizing the variation in predicted SCO response probabilities with component mass fraction， the framework provides decision support for candidate formulation screening and iterative validation through experiments.