Paper deep dive
Identifying Confusion Trends in Concept-based XAI for Multi-Label Classification
Haadia Amjad, Ronald Tetzlaff
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/22/2026, 2:54:15 AM
Summary
This paper investigates Concept-based Explainable AI (CXAI) methods, specifically CRP and CRAFT, applied to multi-label classification tasks using VGG16 and ResNet50 models on the MS-COCO dataset. The study analyzes confusion trends by correlating model performance metrics with CXAI metrics like concept distinctiveness and concept error. Key findings indicate that CXAI reveals learning weaknesses, higher concept distinctiveness reduces label confusion, and environmental concepts expose dataset-induced biases.
Entities (10)
Relation Signals (10)
ResNet50 → istrainedon → MS-COCO
confidence 95% · training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset
VGG16 → istrainedon → MS-COCO
confidence 95% · training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset
CRP → istypeof → CXAI
confidence 95% · We apply two CXAI methods, CRP (Concept Relevance Propagation) and CRAFT
CRAFT → istypeof → CXAI
confidence 95% · We apply two CXAI methods, CRP (Concept Relevance Propagation) and CRAFT
ResNet50 → isusedin → Multi-label classification
confidence 95% · training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset
VGG16 → isusedin → Multi-label classification
confidence 95% · training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset
Environmental Concepts → expose → Dataset Bias
confidence 93% · environmental concepts expose dataset-induced biases
Concept Distinctiveness → reduces → Label Confusion
confidence 92% · higher concept distinctiveness reduces label and concept confusion
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Deep Neural Networks (DNNs) deployed in high-risk domains, such as healthcare and autonomous driving, must be not only accurate but also understandable to ensure user trust. In real-world computer vision tasks, these models often operate on complex images containing background noise and are heavily annotated. To make such models explainable, Concept-based Explainable AI (CXAI) methods need to be assessed for their applicability and problem-solving capacity. In this work, we explore CXAI use cases in multi-label classification by training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset (Microsoft Common Objects in Context). We apply two CXAI methods, CRP (Concept Relevance Propagation) and CRAFT (Concept Recursive Activation FacTorization), to generate concept-level explanations and investigate the overall evaluations. Our analysis reveals three key findings: (1) CXAI highlights learning weaknesses in DNNs, (2) higher concept distinctiveness reduces label and concept confusion, and (3) environmental concepts expose dataset-induced biases. Our results demonstrate the potential of CXAI to enhance the understanding of model generalizability and to diagnose bias instigated by the dataset.
Tags
Links
- Source: https://arxiv.org/abs/2608.15731v1
- Canonical: https://arxiv.org/abs/2608.15731v1
Trouble viewing inline? Open PDF directly →
Full Text
30,870 characters extracted from source content.
Expand or collapse full text
Identifying Confusion Trends in Concept-based XAI for Multi-Label Classification Haadia Amjad Chair of Fundamentals of Electrical Engineering TUD Dresden University of Technology Dresden, Germany e-mail: haadia.amjad@tu-dresden.de Steffen Seitz Chair of Fundamentals of Electrical Engineering TUD Dresden University of Technology Dresden, Germany e-mail: steffen.seitz@tu-dresden.de Ronald Tetzlaff Chair of Fundamentals of Electrical Engineering TUD Dresden University of Technology Dresden, Germany e-mail: ronald.tetzlaff@tu-dresden.de Abstract—Deep Neural Networks (DNNs) deployed in high-risk domains, such as healthcare and autonomous driving, must be not only accurate but also understandable to ensure user trust. In real-world computer vision tasks, these models often operate on complex images containing background noise and are heavily annotated. To make such models explainable, Concept-based Explainable AI (CXAI) methods need to be assessed for their applicability and problem-solving capacity. In this work, we explore CXAI use cases in multi-label classification by training two DNNs, VGG16 and ResNet50, on the 20 most annotated labels in the MS-COCO dataset (Microsoft Common Objects in Context). We apply two CXAI methods, CRP (Concept Relevance Propagation) and CRAFT (Concept Recursive Activation FacTorization), to generate concept-level expla- nations and investigate the overall evaluations. Our analysis reveals three key findings: (1) CXAI highlights learning weaknesses in DNNs, (2) higher concept distinctiveness reduces label and concept confusion, and (3) environmental concepts expose dataset-induced biases. Our results demon- strate the potential of CXAI to enhance the understanding of model generalizability and to diagnose bias instigated by the dataset. Keywords-Concept-based XAI; Multi-Label Classification; Concept Distinctiveness. I. INTRODUCTION Deep Neural Network (DNN) [1] performance is crucial for their adoption in real-world applications. However, understanding their decisions is also impor- tant, especially in high-risk domains like autonomous driving and medical diagnosis. Real-world datasets often vary in resolution and object size, with complex scenes Kilian Göller Chair of Fundamentals of Electrical Engineering TUD Dresden University of Technology Dresden, Germany e-mail: kilian.goeller@tu-dresden.de Carsten Knoll Chair of Fundamentals of Electrical Engineering TUD Dresden University of Technology Dresden, Germany e-mail: carsten.knoll@tu-dresden.de including small, clustered, or overlapping objects. Multi- label datasets, where images have multiple annotations, frequently suffer from class imbalance. This can lead to confusion (i.e., errors made in predicting the correct class/data points) between labels and wrong associations. Even high-performing models that exhibit confusion need deeper analysis. Explainable AI (XAI) methods are useful in revealing these learning patterns [2]. XAI provides interpretability for black-box models [2]. Concept-based XAI (CXAI) identifies semantically meaningful features relevant to a class [3], unlike saliency maps, which are harder to interpret in complex scenes [4]. Concepts reflect how a DNN internally repre- sents a class [5]. However, DNNs may learn unintended associations, concept bias or spurious correlations, where background elements influence classification (e.g., as- sociating “fingers” with a pen) [6]. We refer to non- target concepts produced by such bias as “environmental concepts.” CXAI methods often visualize activation maps or focused image regions [7]. These show both target and environmental concepts. Determining whether an envi- ronmental concept is valid requires further analysis. Its presence may reflect dataset bias or mislearning. In this work, we train two state-of-the-art DNNs, ResNet50 and VGG-16, on the 20 most annotated MS- COCO labels [8]. Using two model checkpoints per architecture, one well-performing and one poor, we evaluate their predictions using CXAI methods: CRP [9] and CRAFT [10]. These methods produce focused region 1Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability visualizations and scores that determine a concept’s contribution to the overall learning (concept importance) or target label learning (concept relevance) of the DNN model. We compare results using concept error and distinctiveness (see Section I, D) to study confusion trends across models. The main contributions of this paper can be summarized as follows: • We demonstrate that CXAI methods can reveal learning weaknesses in deep neural networks. • We find that greater concept distinctiveness is asso- ciated with reduced confusion in label predictions and concept attributions. • We show that environmental concepts can expose dataset-induced biases in model learning and inter- pretation. The remainder of this paper is organized as follows: Section I reviews related studies. Section I describes the experimental setup, including the dataset, DNN mod- els, CXAI methods, and key terminology. Section IV presents the results structured around our three main contributions. Finally, Section V concludes the paper and discusses directions for future work. I. RELATED WORK Various CXAI methods are available for use today, and it is a growing research field. Lee et al. [11] detail the current state of CXAI methods. Their study identifies three main directions for future research: the choice of concepts to explain, the selection of concept representation, and methods to control concepts. Some studies focus on using concepts to detect poten- tial biases in DNN models. Their evaluation emphasizes the relationship between different concepts and classes and aims to expose potential biases in the learning of the DNN. Singh et al. [12] study model biases in both, the model learning process and the model’s semantic understanding (concept biases), by evaluating the DNN model’s ability to recognize a class in the presence and absence of the established context (via learning) for a multi-label classification task. With newer emerging methods in the realm of CXAI, the desire to fully understand how they can be effec- tively used with AI systems increases. The dataset, for example, is an important factor contributing to the mean- ingfulness of the explainability method. The evaluation of CXAI by Ramaswamy et al. [13] addresses important considerations for CXAI methods that influence their effective usage. They emphasize that the impact of the choice of the dataset, even with slight variations in the dataset options, changes the model decision and the explanation provided by a CXAI method. To study the relationship between confusion and concept-based explanations, we select two CXAI meth- ods to answer the “where” (..the important information is) and “what” (..is the important information) questions. CRP, proposed by Achtibat et al. [9], is based on the Layer-wise Relevance Propagation (LRP) method [14]. CRP addresses “what” and “where” explanations by exploiting concepts in hidden layers of a DNN model and locating them in the input data. It assesses the contribution of each concept for a target class; in other words, it introduces concept relevance. CRP utilizes relevance maximization to tune its visualization, which depicts a series of focused concepts. CRAFT is another “what” and “where” method proposed by Fel et al. [10] based on the Grad-CAM method [15]. They utilize Sobol indices to estimate the importance of concepts that have been identified using Non-Negative Matrix Factorization (NMF) recursively, generating sub-concepts (concepts of smaller, more focused areas in the image). Existing research has advanced CXAI by defining concepts and applying them to detect biases and assess dataset effects. Building on this foundation, our work investigates how confusion interacts with concept-based explanations through the lens of CRP and CRAFT. I. EXPERIMENTAL SETUP Just as with any other explainable AI pipeline, our experimentation contains the training of DNNs model and its evaluations and the usage of an XAI method and its evaluation, illustrated in Figure 1. This section contains details of our workflow. Figure 1. Schematic diagram of our experimental setup. A. Dataset MS-COCO [8] is a large-scale dataset widely used for computer vision tasks such as object detection, cap- tioning, segmentation and classification. The 2017 object detection subset includes 80 “things” classes, objects with clear boundaries, across 118,000 images. As test labels are unavailable, we split the training set 90/10, resulting in 106,200 training and 11,800 test images. For our experiments, we focus on the 20 most frequently annotated labels in the training set to ensure sufficient data per class and meaningful inter label relationships. 2Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability i C C B. DNN Models ResNet50 [16] is a popular image classification model due to its residual learning feature, which mitigates information loss. It balances accuracy and efficiency well, and its ImageNet-pretrained weights are widely used [17]. VGG-16 [17], known for its simple and uniform struc- ture of stacked convolutional and fully connected layers, is often used as a baseline for deep learning applications. Despite its larger parameter count, it performs well on classification tasks and is easy to implement. These two models are chosen as they are widely used, and many XAI methods have been proven to work with them. Some of the latest models require large adaptations of XAI methods to be made [18]. Our study focuses on base-level use cases, to be adaptable across different domains; hence, we train ResNet50 and VGG-16 models, pretrained on ImageNetV2 [19], using PyTorch for 350 epochs, saving all checkpoints. For each model, we select two “scenarios” from the saved checkpoints: • Scenario 1 (well-performing model): While both methods offer different perspectives, we do not compare them directly or suggest one is superior. Instead, we use their outputs to explore how label confusion is reflected in learned concepts. We compute concept distinctiveness [21] and concept error [22] for both methods. Concept error is evaluated against a subjective ground truth (detailed in the next section). Additionally, we adapt mutual information to measure shared information between concepts and com- pare these findings to our DNN evaluations to support our hypotheses. D. Explanation of terms (in brief) This sub-section briefly explains some terminologies in CXAI and our adaptations. 1) Concept Distinctiveness: Concept distinctiveness, defined in Eq. (1), measures how unique a concept is compared to others, with values ranging from 0 to 1. Low distinctiveness suggests overlapping or redundant concepts, which may indicate learning errors [21]. – ResNet50: Accuracy: 82.85%, Recall: 85.50, Precision: 58.84, F1 Score: 60.84 v C i · v C j D(C i , C j ) = 1 − |v ||v | (1) – VGG-16: Accuracy: 84.26%, Recall: 86.91, Pre- cision: 59.74, F1 Score: 58.84 • Scenario 2 (poor-performing model): – ResNet50: Accuracy: 58.24%, Recall: 77.04, Precision: 53.82, F1 Score: 42.92 – VGG-16: Accuracy: 52.85%, Recall: 74.50, Pre- cision: 53.62, F1 Score: 46.12 These scenarios are created to have two different sets of performance metrics against which to evaluate explainability. We evaluate models using accuracy, re- call, precision, F1 score, and confusion matrices tailored for multi-label tasks. Specifically, we use the multi- label confusion tensor by Krstinic ́ et al. [20], which accounts for label imbalance—well-suited for the MS- COCO dataset. We also compute Mutual Information (MI) and Jac- card Similarity Coefficient (JSC) between labels. We use these metrics to understand which target labels are more likely to share information or similarities with which predicted labels. C. CXAI Methods We investigate the effect of confusion on two CXAI methods, CRAFT and CRP, across all four model sce- narios. • CRAFT outputs concept importance, representing the overall contribution of each concept to the model’s learning process. • CRP provides concept relevance, indicating the contribution of a concept to specific target classes. Here, v C i and v C j are the concept vectors for concepts C i and C j , respectively. Concept vectors are directions in activation space that capture distinct features [23]. 2) Concept Error: Concept error captures incorrect or irrelevant concept usage during prediction [22]. To ap- proximate accuracy (in binary classification), we define a rough “ground truth” by selecting only those concepts that belong to the target class, excluding environmental concepts. This approach offers an estimate of model con- fusion, though a structured human study is recommended for practical validation. 3) Mutual Information: Mutual information (MI) quantifies the dependency between two variables. In multi-label classification, it measures how much infor- mation one label provides about another. Applied to concepts, MI reflects how much information is shared between two concept vectors, revealing potential depen- dencies or redundancies in learned features [24]. IV. RESULTS In this section, we present our findings based on case studies of different label evaluations. These case studies comprise comparisons of the evaluations described in the previous section. A. Confusion in Labels Can Be Understood by Their Explanations j 3Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability TABLE I. TOP CONFUSION AND MUTUAL INFORMATION SCORES IN SCENARIO 1 OF RESNET50 Class Name Top Confusion Class Top MI Class Jaccard Similarity 1st Score 2nd Score 1st MI Score 2nd MI Score person car 1148.00 chair 1081.70 handbag 0.0221 backpack 0.0176 0.6008 car truck 207.17 bench 173.16 truck 0.0400 traffic light 0.0280 0.1894 motorcycle truck 86.33 handbag 85.70 car 0.0094 person 0.0037 0.1474 truck airplane 118.35 car 117.22 car 0.0399 boat 0.0020 0.0077 boat Parking meter 89.70 car 76.30 chair 0.0682 fork 0.0017 0.0068 TABLE I. TOP CONFUSION AND MUTUAL INFORMATION SCORES IN SCENARIO 2 OF RESNET50 Class Name Top Confusion Class Top MI Class Jaccard Similarity 1st Score 2nd Score 1st MI Score 2nd MI Score person backpack 1995.20 bench 1922.50 tie 0.0221 umbrella 0.0176 0.5470 car backpack 340.70 bench 334.80 boat 0.0399 stop sign 0.0280 0.1159 motorcycle backpack 277.60 handbag 273.41 bicycle 0.0372 car 0.0199 0.1289 truck backpack 209.07 bench 200.09 motorcycle 0.0399 Fire hydrant 0.0077 0.0755 boat car 134.04 bird 133.66 fork 0.0017 refrigerator 0.0010 0.0440 TABLE I. PERCENTAGE OF CO-OCCURRENCE OF TARGET LABEL WITH OTHER LABELS (TOP 20 FREQUENTLY ANNOTATED LABELS) Class Name 1st Class % 2nd Class % person car 13.29 backpack 7.85 car person 69.54 backpack 8.43 motorcy -cle person 79.55 car 39.32 truck person 65.15 car 59.80 boat person 65.69 car 8.66 traffic light car 61.22 person 59.19 bench person 73.75 car 14.63 bird person 24.56 boat 7.29 sheep person 24.07 dog 7.59 backpack person 91.06 car 18.69 umbrella person 86.87 handbag 28.81 handbag person 90.95 backpack 24.62 kite person 92.84 car 11.54 bottle person 53.65 cup 34.65 cup person 52.76 dining table 50.92 bowl dining table 47.76 person 40.73 banana person 41.37 bowl 23.05 potted plant person 44.07 chair 38.61 dining table person 49.58 chair 43.29 book dining table 75.61 cup 52.97 TABLE IV. CXAI METHOD EVALUATION COMPARED WITH CONFUSION SCORE FOR ’PERSON’ LABEL Label Model CXAI Method Concept Error Concept Distinct -iveness Confus -ion Score Person ResNet Scenario 1 Craft 0.20 0.76 0.09 Person ResNet Scenario 2 Craft 0.38 0.48 0.26 Person VGG-16 Scenario 1 Craft 0.24 0.71 0.12 Person VGG-16 Scenario 2 Craft 0.41 0.43 0.28 Label confusion occurs when models struggle to dis- tinguish between classes with overlapping features or co- occurring contexts, often due to ambiguous data, mis- labeling, or internal misinterpretation. We hypothesize that CXAI methods, particularly through MI and concept distinctiveness, can reveal whether confusion stems from visual similarity, dataset bias, or how the model encodes relationships between labels. Tables I and I present confusion and MI scores for three highly confused classes across both ResNet50 scenarios. In scenario 1, person is confused with car and chair, while car overlaps with truck and bench. MI analysis shows that person shares high information content with handbag and backpack, and car with truck and traffic light. These associations indicate that the model is not learning isolated class-specific features, but instead forming dependencies based on recurring visual or contextual co-occurrence. Table I supports this, showing frequent joint appearance of labels such as person and accessories, or car and truck, which reinforces these spurious links. Table IV further highlights the role of CXAI metrics 4Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability in understanding confusion. In scenario 1, where models perform better, person has lower concept error and higher distinctiveness, aligning with reduced confusion. In scenario 2, we observe the opposite: increased concept error, lower distinctiveness, and significantly higher con- fusion scores. These patterns suggest that when a model lacks distinct conceptual boundaries between classes, it tends to rely more heavily on misleading contextual aspects. Together, these findings show how CXAI methods help expose the roots of confusion. By combining ex- planations with performance metrics and co-occurrence statistics, we gain a clearer view of when confusion reflects real-world visual similarity versus when it results from dataset bias or poor internal representations. B. Distinctiveness Reduces Conceptual Confusion When a concept is distinct, its features are unique and specific, allowing it to be more accurately defined and recognized. In contrast, concepts derived from confused or overlapping labels tend to be “confused” themselves, as they learn features that are shared across multiple classes rather than those unique to their true class. This issue arises from concept bias, where the model may associate a class with irrelevant features that co-occur with other classes, as shown in Figure 2. Figure 2. Concepts of class “tennis racket” in scenario 1 of VGG-16. We can see that “person” is heavily present in these explanations. TABLE V. MUTUAL INFORMATION, CONCEPT DISTINCTIVENESS, AND CONCEPT ERROR IN SCENARIO 1 OF RESNET50 Class Name Top MI (Concept) Lowest Distinctive (CRP) Lowest Distinctive (CRAFT) Concept Error 1st 2nd 1st 2nd 1st 2nd Value person car backpack car backpack car tennis racket 0.7291 car truck bus truck traffic light handbag truck 0.5385 dining table chair cup chair fork person chair 0.0166 From the information given in Table VI, it is evident that a poor-performing model is not ideal for concept- based explanations due to the lack of clear distinctions between classes. This can be seen in scenario 2 of ResNet50, where classes like person show less dis- tinctiveness with other unrelated classes. In scenario 1, shown in Table V, we see a more effective distinction be- tween highly confused classes like car and person, which TABLE VI. MUTUAL INFORMATION, CONCEPT DISTINCTIVENESS, AND CONCEPT ERROR IN SCENARIO 2 OF RESNET50 Class Name Top MI (Concept) Lowest Distinctive (CRP) Lowest Distinctive (CRAFT) Concept Error 1st 2nd 1st 2nd 1st 2nd Value person car tennis racket backpack bottle backpack umbrella 0.8136 car truck traffic light bench fire hydrant backpack boat 0.6388 dining table cup bottle chair fork person potted plant 0.0753 indicates that a well-performing model actively tries to separate these difficult-to-distinguish classes (previously established based on confusion scores, see Table I, V, VI and I). By focusing on distinctiveness metrics and correlat- ing them with confusion patterns in Table I and co- occurrence in Table I, we see that increasing concept distinctiveness can significantly aid in or point to im- proved model performance. This insight not only helps in diagnosing where models are struggling but also guides how to curate datasets and improve feature learning to reduce confusion and improve overall classification accuracy. C. Environmental Concepts Reveal Dataset Biases Figure 3. Environmental concepts generated from CRP for class “car” in scenario 1 and 2 of ResNet50. TABLE VII. MUTUAL INFORMATION (CONCEPT), MUTUAL INFORMATION AND CONFUSION SCORES IN SCENARIO 1 OF VGG-16 Class Name Top MI (Concept) Top MI (Class) Top Con fusion 1st 2nd 1st 2nd 1st 2nd umbrella person handbag backpack handbag person car dining table chair fork chair cup apple person traffic light person car car fire hydrant person car Environmental concepts emerge from concept bias and often reflect patterns in the training dataset. We ob- serve that classes within the same “supercategory” (e.g., sports: baseball glove, tennis racket) tend to produce biased explanations, frequently including environmen- tal concepts from related classes, illustrated in Figure 3. This suggests that, beyond model performance, the diversity and distinctiveness of training samples play a key role in learning meaningful class representations. 5Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability Figure 4. Concepts generated by CRP on OSDaR23 dataset for class "person". Table VII illustrates the top mutual information and confusion scores for selected classes. For instance, din- ing table in scenario 1 is frequently associated with chair, person, apple, and cup, which are labels that share semantic but not structural similarity. Such associations, while intuitive to humans, suggest that the model is not generalizing but instead relying on frequent co- occurrences, which is problematic in deployed systems. High concept error rates for classes like umbrella, per- son, handbag, and car, paired with low distinctiveness scores between semantically unrelated objects (e.g., um- brella and traffic light), reinforce this concern, especially when models perform poorly. To further support this, we evaluate OSDaR23 [25], a multi-sensor dataset for autonomous train driving. Despite strong accuracy (95.92%) and F1 (79.93) on a ResNet50 model trained on its RGB subset, CXAI explanations reveal low generalizability. Since person consistently appears near platforms or staircases, CRP visualizations heavily rely on these backgrounds, none of which are labeled in the dataset, as illustrated in Figure 4. As a result, person has the lowest distinctiveness score with track, and a high concept error, indicating dangerous misattribution. These findings highlight how environmental concepts reveal dataset-induced biases that compromise gener- alization. In real-world or high-risk applications, such as autonomous systems, these misleading correlations can reduce model reliability. Diverse and well-annotated datasets are essential to prevent concept bias and ensure models learn robust, semantically accurate representa- tions. V. CONCLUSION AND FUTURE WORK Our study demonstrates that confusion in multi-label classification is directly reflected in concept-based ex- planations. By comparing model evaluations with CXAI properties, we observe that label confusion often results from overlapping or spurious environmental concepts, emphasizing the role of CXAI in uncovering learning biases and assessing model generalizability. We further show that concept distinctiveness is inversely related to conceptual confusion, models with higher distinctive- ness show clearer feature boundaries and reduced bias, while lower distinctiveness leads to shared or incorrect associations across classes. CRP and CRAFT help iden- tify such conceptual ambiguities, making them useful tools for model diagnosis. Finally, our results highlight that environmental concepts can reveal dataset-induced biases, especially in cases where co-occurring objects affect model learning. In datasets with label imbalance or strong contextual patterns, models may form misleading correlations, reducing their ability to generalize. This is particularly problematic in high-risk applications, rein- forcing the need for diverse, well-annotated datasets to ensure robust and reliable AI models. For future work, this case study can be extended to more complex models and datasets. ACKNOWLEDGEMENT This work is partly supported by BMFTR (Federal Ministry of Research, Technology and Space) in DAAD project 57616814 (SECAI, School of Embedded Com- posite AI, https://secai.org/). REFERENCES [1] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, p. 436–444, 2015. [2] W. Samek, G. Montavon, S. Lapuschkin, C. J. Anders, and K.-R. Müller, “Explaining deep neural networks and beyond: A review of methods and applications,” Proceedings of the IEEE, vol. 109, no. 3, p. 247–278, 2021. [3] E. Poeta, G. Ciravegna, E. Pastor, T. Cerquitelli, and E. Baralis, Concept-based explainable artificial intelli- gence: A survey, Preprint arXiv:2312.12936, 2023. [4] R. Müller, M. Thoß, J. Ullrich, S. Seitz, and C. Knoll, “Interpretability is in the eye of the beholder: Human versus artificial classification of image segments gen- erated by humans versus xai,” International Journal of Human-Computer Interaction, p. 2371–2393, 2024. [5] S. S. Y. Kim, E. A. Watkins, O. Russakovsky, R. Fong, and A. Monroy-Hernández, “"help me help the ai": Understanding how explainability can support human- ai interaction,” in Proceedings of the 2023 CHI Confer- ence on Human Factors in Computing Systems, 2023, p. 1–17. [6] P. Stock and M. Cisse, “Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering bi- ases,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, p. 504–519. 6Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability [7] B. Kim et al., “Interpretability beyond feature attribu- tion: Quantitative testing with concept activation vectors (tcav),” in International Conference on Machine Learn- ing, PMLR, 2018, p. 2668–2677. [8] T.-Y. Lin et al., “Microsoft coco: Common objects in context,” in European Conference on Computer Vision, Springer, 2014, p. 740–755. [9] R. Achtibat et al., “From attribution maps to human- understandable explanations through concept relevance propagation,” Nature Machine Intelligence, vol. 5, p. 1006–1019, 2023. [10] T. Fel et al., “Craft: Concept recursive activation fac- torization for explainability,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, p. 2711–2721. [11] J. H. Lee, G. Mikriukov, G. Schwalbe, S. Wermter, and D. Wolter, “Concept-based explanations in computer vision: Where are we and where could we go?” In European Conference on Computer Vision, Springer, 2025, p. 266–287. [12] K. K. Singh et al., “Don’t judge an object by its context: Learning to overcome contextual bias,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, p. 11 070–11 078. [13] V. V. Ramaswamy, S. S. Y. Kim, R. Fong, and O. Rus- sakovsky, “Overlooked factors in concept-based expla- nations: Dataset choice, concept learnability, and human capability,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, p. 10 932–10 941. [14] A. Binder, G. Montavon, S. Lapuschkin, K.-R. Müller, and W. Samek, “Layer-wise relevance propagation for neural networks with local renormalization layers,” in International Conference on Artificial Neural Networks, Springer International Publishing, 2016, p. 63–71. [15] R. R. Selvaraju et al., “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision, 2017, p. 618–626. [16] K. He, X. Zhang, S. Ren, and J. Sun, “Deep resid- ual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, p. 770–778. [17] K. Simonyan and A. Zisserman, “Very deep convolu- tional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. [18] H. Ibrahim Aysel, X. Cai, and A. Prugel-Bennett, “Explainable artificial intelligence: Advancements and limitations,” Applied Sciences, vol. 15, p. 7261, 2025. [19] J. Deng et al., “Imagenet: A large-scale hierarchical im- age database,” in IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2009, p. 248–255. [20] D. Krstinic ́, A. K. Skelin, I. Slapnicˇar, and M. Braovic ́, “Multi-label confusion tensor,” IEEE Access, vol. 12, p. 9860–9870, 2024. [21] B. Wang, L. Li, Y. Nakashima, and H. Nagahara, “Learning bottleneck concepts in image classification,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2023, p. 10 962– 10 971. [22] P. W. Koh et al., “Concept bottleneck models,” in International Conference on Machine Learning, PMLR, 2020, p. 5338–5348. [23] L. O’Mahony, V. Andrearczyk, H. Müller, and M. Graziani, “Disentangling neuron representations with concept vectors,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2023, p. 3770–3775. [24] M. E. Zarlenga et al., Concept embedding models: Beyond the accuracy-explainability trade-off, Preprint arXiv:2209.09056, 2022. [25] R. Tagiew et al., “Osdar23: Open sensor data for rail 2023,” in International Conference on Robotics and Automation Engineering, 2023, p. 270–276. 7Copyright (c) IARIA, 2025. ISBN: 978-1-68558-318-7 Courtesy of IARIA Board and IARIA Press. Original source: ThinkMind Digital Library https://w.thinkmind.org EXPLAINABILITY 2025 : The Second International Conference on Systems Explainability