Credal Classification through an Ensemble of Confidence-Aware TabTransformers and its Application to Fraud Detection

We address a multi-class classification problem on a tabular dataset comprising both continuous and categorical features, as commonly found in financial and actuarial domains. To achieve robust and reliable predictions, we propose a credal classifier built from an ensemble of augmented probabilistic models, each capable of self-assessing its confidence in the predicted class distribution for a given instance. Each base learner in the ensemble is implemented as a Confidence-Aware TabTransformer: a neural architecture that combines a sequence of transformer layers to process categorical variables, followed by a multi-layer perceptron for classification. In parallel, a dedicated confidence head outputs the model’s self-assessed reliability score. For ensemble aggregation, we explore multiple confidence-aware combination rules, applied after a [Formula: see text]-quantile filtering step to remove low-discordant outlier distribution. We demonstrate the effectiveness of this methodology on a fraud detection task, which is a representative binary classification problem with severe class imbalance.