Conformal deep forest for uncertainty-aware classification

Uncertainty in deep learning models significantly impacts their performance, robustness, and reliability, making explicit uncertainty quantification a critical research focus. However, existing methods often fail to incorporate relationships between classes into uncertainty quantification, which are essential for capturing class-dependent uncertainty. To address this challenge, this paper presents two key contributions. First, the conformal dual uncertainty metric (CDUM), derived from conformal prediction, is proposed as a novel uncertainty criterion. CDUM quantifies class-wise uncertainty by capturing class relationships from two perspectives: the influence of other classes on the current class and the dominance of the current class over others. Second, the conformal deep forest (CDForest), an uncertainty-aware deep model, is proposed by incorporating three CDUM-based modules into the deep forest (DF) architecture. Leveraging these CDUM-based modules, CDForest achieves three core capabilities: (1) layer-wise uncertainty filtering to reduce high-uncertainty features, (2) class-wise weighted averaging to refine ensemble predictions, and (3) class-wise aggregation for conformal predictors to highlight class-specific differences and facilitate more precise uncertainty quantification. Extensive experiments demonstrate that CDForest significantly outperforms state-of-the-art models in terms of classification accuracy, robustness against noise, and decision reliability, further confirming its effectiveness as a highly competitive solution for uncertainty-aware classification tasks.