Enhancing early-stage breakthrough technology identification: A CatBoost model optimized with the SABO algorithm

Early identification of breakthrough technologies is essential for strategic R&D planning, but it remains a major challenge in technology analysis and strategic management. In their early stages, these technologies often appear in patent data with weak and ambiguous indicators and are greatly outnumbered by incremental innovations. This results in low feature discriminability and severe class imbalance. To address these challenges, we present a new automated framework that combines CatBoost with the SABO heuristic optimization algorithm. In this approach, CatBoost's key hyperparameters are treated as SABO search agents and are optimized to improve the minority-class F1-score using stratified cross-validation. This approach directs the model to achieve high sensitivity and robustness, given the imbalanced data and subtle indicators. Experimental validation using the Derwent Innovation patent data set shows that SABO-tuned CatBoost significantly outperforms both standard CatBoost and other optimizer-based models in F1-score for breakthrough identification. Model interpretation through feature importance and SHAP analysis identifies discriminative indicators for early-stage breakthroughs. The proposed framework offers an effective and interpretable tool for technology forecasting under these constraints, supporting early opportunity discovery and data-driven innovation strategy.