Gaussian evidential ensemble learning of logistic classifiers
Liping Liu · Knowledge-Based Systems · 2025
• Gaussian ensemble method is widely applicable to any learned models based on MLE • The Gaussian method allows direct combination of models rather than model predictions • A Gaussian ensemble keeps memory of its members, allowing for sequential updating • The Gaussian method performs better than the reference model in experiments Leveraging the asymptotic normality of maximum likelihood estimators, this paper proposes a novel ensemble learning method based on Gaussian belief functions. Unlike traditional ensemble approaches that combine model predictions, the proposed method directly fuses the learned models, enabling distributed training and sequential learning without revisiting past ensemble members or relying on their ordering or weighting. This approach is particularly well-suited for large-scale or streaming applications, where retraining or retaining all prior models is impractical. The method is broadly applicable to supervised, unsupervised, and semi-supervised settings, provided the base learners are trained via maximum likelihood estimation. To demonstrate its utility, the method is applied to ensembles of logistic classifiers and evaluated using both simulated datasets and a real-world dataset from the UCI Machine Learning Repository. Experimental results show that the Gaussian evidential ensemble consistently outperforms its constituent members, achieving predictive performance comparable to or better than a reference model trained on fully aggregated data. Moreover, the method exhibits desirable asymptotic behavior and robustness in sequential learning, affirming its theoretical soundness and practical relevance for scalable ensemble learning. (keywords: Ensemble Learning, Evidential Ensemble, Logistic Classifiers, Dempster’s Rule, Gaussian Belief Functions)