SACX: Suppression–Absence–Competition eXplanations for the why-not question in visual recognition

Introduction Existing attribution-based explainable artificial intelligence (XAI) methods for convolutional neural network (CNN) classifiers primarily address why a particular class was predicted, while providing limited insight into why a competing class was not selected. To address the question “Why class A and not class B?”, we introduce Suppression-Absence-Competition eXplanations (SACX), a framework for structured contrastive visual explanations. Methods SACX characterizes the contrastive margin z a − z b using three spatially aligned heatmaps: H comp , representing positive evidence for the predicted class; H sup , representing negative attribution for the competing foil class; and H miss , identifying regions where expected foil-class features are absent through prototype-based similarity. We introduce three direction-aware faithfulness metrics: suppression-removal AUC (SR-AUC), feature-injection AUC (FI-AUC), and localization fidelity under injection (LFI@k). We additionally evaluate suppression using a fixed-budget placement test (SR@k). Experiments are conducted on CUB-200-2011 and Pattern Analysis, Statistical Modelling and Computational Learning (PASCAL) Visual Object Classes (VOC) 2007 using fine-tuned ResNet-50 models and a frozen DINOv2 ViT-B/14 backbone with a linear probe. Results On fine-tuned ResNet-50, injecting features at top-ranked H miss locations outperforms random placement, improving the foil-class logit by Δ z b = +0.41 (95% CI: 0.39–0.42) on CUB and +1.29 (95% CI: 1.23–1.36) on VOC. On the frozen DINOv2 backbone, all injection variants reduce the foil-class logit, but top-ranked H miss locations remain the least disruptive, achieving an LFI@20% top-minus-random gap of +0.68. The LFI gaps remain stable across three K-means and three probe seeds (standard deviation ≤ 0.04). The SR@k evaluation shows that H sup outperforms random removal at tight budgets (+0.09 at k = 5% on CUB Top-2), although suppression effects are generally weak for false-negative cases. Discussion SACX provides an interpretable structural decomposition of the existing contrastive attribution signal rather than introducing a new aggregate map, as H contra is mathematically equivalent to Contrastive Grad-CAM. Its prototype-based missing-evidence component preserves spatial localization rankings even under frozen self-supervised representations, although absolute injection gains require task-specific supervision. The suppression component is interpreted as negative attribution rather than a guarantee of causal suppression. These findings demonstrate the value of separating competition, suppression, and missing evidence when explaining why a competing class was not predicted.