Identifying Confusion Trends in Concept-based XAI for Multi-Label Classification

• The study evaluates CRP and CRAFT on ResNet50 and VGG-16 models trained on the 20 most annotated MS-COCO labels.It compares well-performing and poor checkpoints for each architecture using concept error and distinctiveness to study confusion trends.Source: §I · pp. 1–2
• CXAI methods reveal learning weaknesses in deep neural networks.Source: §I · p. 2
• Greater concept distinctiveness is associated with reduced confusion in label predictions and concept attributions.Source: §I · p. 2
• Environmental concepts can expose dataset-induced biases in model learning and interpretation.Source: §I · p. 2

II. RELATED WORK

• Lee et al. identify concept choice, concept representation, and concept control as three directions for future CXAI research.Source: §II · p. 2
• Singh et al. examine model and concept biases by testing class recognition with and without its established context in multi-label classification.Source: §II · p. 2
• Ramaswamy et al. show that even slight dataset-option variations can change model decisions and the explanations produced by CXAI methods.Source: §II · p. 2
• CRP explains what and where by locating hidden-layer concepts and assessing their relevance to a target class, while CRAFT recursively identifies sub-concepts and estimates their importance.CRP uses relevance maximization; CRAFT uses Sobol indices and Non-Negative Matrix Factorization.Source: §II · p. 2
• Building on research that defines concepts and uses them to detect bias and assess dataset effects, this work investigates confusion through CRP and CRAFT.Source: §II · p. 2

III. EXPERIMENTAL SETUP

• The workflow trains and evaluates DNNs, then applies and evaluates an XAI method, as illustrated in Figure 1.Source: §III · p. 2, Figure 1 · p. 2
• MS-COCO contributes 106,200 training images and 11,800 test images after a 90/10 split, with experiments focused on its 20 most frequently annotated labels.The dataset’s 2017 object-detection subset contains 80 “things” classes across 118,000 images; test labels were unavailable, so the training set was split instead.Source: §III · p. 2

B. DNN Models

• ResNet50 and VGG-16 [17] were selected as widely used models compatible with XAI methods, and trained from ImageNetV2-pretrained weights for 350 epochs.The study targets base-level use cases to support adaptation across domains; all checkpoints were saved. ResNet50 is described as using residual learning, while VGG-16 has a simple, uniform structure.Source: B. DNN Models · p. 3
• Scenario 1 represents well-performing checkpoints for both models.ResNet50: Accuracy 82.85%, Recall 85.50, Precision 58.84, F1 Score 60.84. VGG-16: Accuracy 84.26%, Recall 86.91, Precision 59.74, F1 Score 58.84.Source: B. DNN Models · p. 3
• Scenario 2 represents poor-performing checkpoints for both models.ResNet50: Accuracy 58.24%, Recall 77.04, Precision 53.82, F1 Score 42.92. VGG-16: Accuracy 52.85%, Recall 74.50, Precision 53.62, F1 Score 46.12.Source: B. DNN Models · p. 3
• Models are evaluated using accuracy, recall, precision, F1 score, and multi-label confusion matrices tailored to account for label imbalance in MS-COCO.Source: B. DNN Models · p. 3
• Mutual Information and Jaccard Similarity Coefficient are computed between labels to identify likely information-sharing or similarity relationships between target and predicted labels.Source: B. DNN Models · p. 3

• The analysis uses CRAFT and CRP across all four model scenarios to examine how label confusion is reflected in learned concepts, without ranking the methods.Source: C. CXAI Methods · p. 3
• CRAFT reports each concept’s overall importance to model learning, whereas CRP reports concept relevance to specific target classes.Source: C. CXAI Methods · p. 3
• For both methods, the study computes concept distinctiveness and concept error, evaluates error against a subjective ground truth, and adapts mutual information to measure shared information between concepts.These findings are compared with DNN evaluations to support the study’s hypotheses.Source: C. CXAI Methods · p. 3

D. Explanation of terms (in brief)

• The subsection briefly explains CXAI terminology and the authors’ adaptations.Source: D. Explanation of terms (in brief) · p. 3
• Concept distinctiveness measures how unique a concept is compared with others, on a scale from 0 to 1.It is defined in Eq. (1).Source: 1) Concept Distinctiveness · p. 3
• Low concept distinctiveness suggests overlapping or redundant concepts and may indicate learning errors.Source: 1) Concept Distinctiveness · p. 3

1) Concept Distinctiveness:

• Concept error captures incorrect or irrelevant concept use, and concept vectors represent distinct features as directions in activation space.The target-only concept reference is an estimate; practical validation with a structured human study is recommended.Source: 2) Concept Error · p. 3, 1) Concept Distinctiveness · p. 3

3) Mutual Information:

• Mutual information measures how much one label informs another, and between concept vectors it reveals shared information, dependencies, or redundancies in learned features.Source: 3) Mutual Information · p. 3

IV. RESULTS

• The results are presented through case studies comparing the evaluations described in the preceding section.Source: §IV · p. 3

A. Confusion in Labels Can Be Understood by Their Explanations

• The case-study tables cover top confusion and mutual-information scores for two ResNet50 scenarios, label co-occurrence, and CXAI evaluation against person confusion.Source: §IV · p. 3, Table I · p. 4, Table II · p. 4
• In scenario 1, person had lower concept error and higher distinctiveness with reduced confusion; scenario 2 showed the opposite pattern and greater reliance on misleading context.The passages characterize scenario 1 as one where models perform better, and scenario 2 as having significantly higher confusion scores.Source: A. Confusion in Labels Can Be… · p. 5
• Combining explanations with performance metrics and co-occurrence statistics clarifies whether confusion reflects visual similarity, dataset bias, or poor internal representations.Source: A. Confusion in Labels Can Be… · p. 5

B. Distinctiveness Reduces Conceptual Confusion

• Overlapping labels can produce confused concepts that learn shared rather than class-specific features; tennis-racket explanations heavily include person.The passage attributes this concept bias to irrelevant features that co-occur with other classes.Source: B. Distinctiveness Reduces Conceptual… · p. 5, Figure 2 · p. 5
• ResNet50’s scenario 1 shows more effective distinction between confused classes such as car and person than scenario 2, where person is less distinct from unrelated classes.The findings suggest that a better-performing model actively separates difficult-to-distinguish classes.Source: Table V · p. 5, Table VI · p. 5, B. Distinctiveness Reduces Conceptual… · p. 5
• Correlating distinctiveness with confusion and co-occurrence patterns can help diagnose model weaknesses and guide dataset curation and feature learning to reduce confusion and improve classification accuracy.The passage presents increased concept distinctiveness as a way to aid or indicate improved model performance.Source: B. Distinctiveness Reduces Conceptual… · p. 5

C. Environmental Concepts Reveal Dataset Biases

• Classes within the same supercategory often produce biased explanations with environmental concepts from related classes, suggesting training-sample diversity and distinctiveness matter for meaningful representations.The paper illustrates this pattern for sports classes such as baseball glove and tennis racket.Source: C. Environmental Concepts Reveal… · p. 5, Figure 3 · p. 5
• Dining table in scenario 1 was associated with chair, person, apple, and cup, while high concept errors and low distinctiveness for other classes indicate reliance on frequent cooccurrences rather than generalization.The passage notes low distinctiveness between semantically unrelated umbrella and traffic light, especially when model performance is poor.Source: Table VII · p. 5, C. Environmental Concepts Reveal… · p. 6
• Despite 95.92% accuracy and 79.93 F1 on OSDaR23 RGB data, ResNet50 relied on unlabeled platforms and staircases for person, which had its lowest distinctiveness with track and high concept error.CRP visualizations revealed this background reliance despite strong performance, indicating low generalizability and dangerous misattribution.Source: C. Environmental Concepts Reveal… · p. 6, Figure 4 · p. 6
• In real-world and high-risk applications, misleading environmental correlations can reduce reliability, so diverse, well-annotated datasets are essential for robust, semantically accurate representations.These findings characterize environmental concepts as revealing dataset-induced biases that compromise generalization.Source: C. Environmental Concepts Reveal… · p. 6

V. CONCLUSION AND FUTURE WORK

• Label confusion is reflected in concept explanations, often through overlapping or spurious environmental concepts that reveal learning biases and inform assessments of model generalizability.The authors note that co-occurring objects and strong contextual patterns can produce misleading correlations that reduce generalization.Source: §V · p. 6
• Higher concept distinctiveness corresponds to less conceptual confusion, clearer feature boundaries, and reduced bias; lower distinctiveness can produce shared or incorrect class associations.CRP and CRAFT help identify these conceptual ambiguities for model diagnosis.Source: §V · p. 6
• Environmental concepts can expose dataset-induced biases, especially when co-occurring objects, label imbalance, or strong contextual patterns encourage misleading correlations.Such correlations can reduce a model’s ability to generalize, a particular concern in high-risk applications.Source: §V · p. 6
• The authors recommend diverse, well-annotated datasets for robust and reliable AI, and propose extending this case study to more complex models and datasets.Source: §V · p. 6