A stacking ensemble learning approach for scientific data classification

Abstract

To address the limitations of individual models in adaptive performance and the challenges of utilizing high-dimensional, sparse text features in automated scientific data classification, this paper proposes an approach based on an improved Stacking ensemble learning framework. Using dataset names and keywords as inputs, the method constructs lightweight text features via TF-IDF, bypassing the need for complex feature engineering. The experimental evaluation was conducted on a real-world scientific dataset comprising 4,505 samples across five disciplines. The performance of the proposed stacking model was benchmarked against four individual models and two simple ensemble methods (soft voting and weighted average). The results indicate that SVC exhibits the best adaptability to high-dimensional, sparse short-text features. Furthermore, ensemble models consistently outperformed individual models. Notably, the effectiveness of the stacking ensemble is determined by the quality and complementarity of the base models rather than their quantity. The proposed Stacking model achieved the highest classification performance, reaching an accuracy of 0.9024 on the test set. This lightweight, efficient, and stable method provides robust technical support for the automated classification and hierarchical governance of scientific data.