关键词: 数据平衡, 集成学习, 颠覆性技术识别, 模型性能优化, 机器学习

Abstract:

[Purpose/Significance] Disruptive technology identification has become an increasingly important research topic in the context of rapid technological evolution and strategic decision-making for governments and enterprises. However, existing data-driven identification approaches often suffer from two critical limitations. First, disruptive technology datasets are typically characterized by severe class imbalance, where truly disruptive cases constitute only a small fraction of the total samples, leading to biased learning and poor generalization. Second, most existing studies rely on a single machine learning model, which limits the ability to capture complex and heterogeneous patterns embedded in high-dimensional technical text features. These issues restrict the robustness, accuracy, and practical applicability of current identification frameworks. To address these challenges, this study aims to construct an optimized disruptive technology identification model that jointly considers data imbalance mitigation and model performance enhancement, thereby improving the reliability and stability of predictive results and contributing to methodological advancements in technology intelligence and innovation management research. [Method/Process] Based on the reproduction of a widely used baseline model built upon XGBoost, this study proposed a two-stage optimization framework integrating data resampling and ensemble learning. In the data preprocessing stage, a hybrid SMOTE-ENN sampling strategy was employed to reconstruct the training dataset. The SMOTE component synthetically generated minority class samples to enhance class representation, while the ENN component removed ambiguous and noisy samples from overlapping regions, thus achieving a balance between noise reduction and information preservation. This strategy effectively alleviated the adverse impact of class imbalance on model learning without excessively distorting the original data distribution. In the modeling stage, a stacking-based ensemble learning framework was constructed by integrating multiple heterogeneous base learners, including XGBoost, LightGBM, Extra Trees, and Support Vector Machines. These base models were selected to capture complementary decision boundaries and feature interactions from different learning perspectives. A Random Forest model was further employed as a meta-learner to aggregate the outputs of the base learners and perform higher-level feature integration. Through this hierarchical learning mechanism, the proposed framework enhanced both representation capability and predictive robustness, enabling more accurate identification of disruptive technologies under complex and noisy data conditions. [Results/Conclusions] Extensive experimental evaluations demonstrate that the proposed optimization model significantly outperforms the baseline XGBoost model across multiple core performance metrics, including Accuracy, Precision, Recall, and F1-Score. Notably, the F1-Score, which is substantially improved from 0.63 to 0.98, indicates a marked enhancement in the model's ability to correctly identify minority disruptive technology samples while maintaining high overall stability. The results confirm that the combined application of hybrid resampling and ensemble learning effectively addresses the challenges of sample imbalance and model bias in disruptive technology identification tasks. In conclusion, the proposed framework provides a robust and scalable solution for identifying disruptive technologies in high-dimensional, imbalanced data scenarios. Beyond improving prediction accuracy, this study offers methodological insights for technical text modeling and innovation analytics. Its approach can be easily adapted to other fields with similar data imbalance and complexity issues. Future research may further explore adaptive sampling strategies and deep learning-based ensemble architectures to enhance temporal and semantic representation capabilities.

Key words: data balance, integrated learning, subversive technology identification, model performance optimization, machine learning