Feature Weight Optimization in Machine Learning Classifiers for Conflict Escalation Early Warning: Evidence from Diplomatic Signals and News Text

Early warning of armed conflict escalation remains a central challenge at the intersection of computational social science and national security analytics. Existing machine learning pipelines for conflict prediction typically treat all input features with equal or heuristically assigned weights, overlooking the differential informativeness of diplomatic signals versus macroeconomic indicators across varying escalation phases. This paper proposes a structured feature weight optimization framework integrating diplomatic statement tone data from GDELT/CAMEO event coding with news-derived LDA topic features. Two baseline classifiers---Random Forest and Gradient Boosting---are compared under standard and optimized weighting conditions. SHAP-based interpretability analysis quantifies the marginal contribution of each feature group to escalation-onset prediction. Experiments on a longitudinal country-month panel (2010--2023, N = 25,074) demonstrate that the proposed weighting strategy improves AUC-ROC by 5.2 percentage points over unweighted baselines while reducing false alarm rates by 11.2%, offering actionable guidance for intelligence analysts prioritizing early escalation indicators across heterogeneous data streams.