Dynamic Model Weighting via Stacking-Meta Learning for Multi-Class Classification

• Agostinho Freire UPE
• João V. R. de Andrade UPE
• Bruno J. T. Fernandes UPE
• Leandro H. de S. Silva UPE / IFPB
• Gabriel Albuquerque UPE
• Mário Guerra UPE
• Igor Balbino UPE
• Monique Suellen Tomaz UPE
• João Fausto L. de Oliveira UPE

Resumo

Ensemble learning has emerged as a powerful solution for addressing the challenges associated with complex, high-dimensional, and imbalanced datasets across structured and unstructured domains. However, traditional ensemble approaches often lack mechanisms for dynamically selecting and weighting models based on their predictive strengths and uncertainties, thus limiting their performance in multi-class classification tasks. To overcome this limitation, we propose a novel Stacking-Meta Learning approach, which transfers the model selection and weighting responsibilities to an adaptive outer meta-model. This meta-model dynamically identifies and leverages statistical patterns within the probabilistic outputs of diverse inner models. Our approach significantly enhances predictive accuracy and robustness by adaptively balancing precision and recall across changing data patterns. We clarify the design of our dynamic scoring function and its properties (boundedness, monotonicity, and convex combination). Extensive experimental evaluation on two representative datasets shows that the Stacking-Meta Learning method achieves competitive or superior performance against strong single models and a bagging ensemble, in both structured and unstructured multi-class settings. In low-data regimes, it maintains robust performance. We discuss computational considerations and outline ablations and additional baselines as directions for future work.

Palavras-chave: Learning systems, Adaptation models, Uncertainty, Limiting, Computational modeling, Predictive models, Probabilistic logic, Robustness, Ensemble learning, Periodic structures