EXPLAINS 2026 Abstracts

Area 1 - Social and Legal Issues

Short Papers

Paper Nr:	131
Title:	Attribution Is Not Accountability: Explainable AI as Compliance Evidence under the EU AI Act and the GDPR
Authors:	Sheyla Leyva-Sánchez and Otto Wagner
Abstract:	Organisations deploying high-risk AI increasingly file a single explainability output, typically a feature-attribution plot, as evidence that a decision is accountable under the EU AI Act and the GDPR. We argue that this equation is mistaken and ethically hazardous: attribution answers only part of what the obligations ask, is not faithful to the model by default, and several widely used methods are manipulable or insensitive to the model they purport to explain. Filing such an output as ``the reason'' for a decision is a form of accountability-washing. We defend an alternative. Accountability evidence requires a set of properties an explanation must have, a layered artifact that jointly achieves them, and an auditable lifecycle that retains and validates it in production. It also requires excluding, on principle, methods that cannot stand as primary compliance evidence. Grounded in EU law and the standards that operationalise it, this position paper reframes the choice of explanatory evidence as an accountability question, not a matter of tool selection.

Paper Nr:	50
Title:	The Myth of the Box: Causation and Comprehensibility in Neural Network Behavior
Authors:	Ian O'Loughlin
Abstract:	Because neural network architectures defy attempts to trace the individual processes responsible for any given, particular output, deep learning systems that rely on these architectures are often characterized as ``black boxes''. The supposition that there are inaccessible features of the internal process that are responsible for individual behaviors deeply informs our understanding of, our trust in, and our experimentation with these systems. This supposition, however, rests on an assumption about the nature of causation in systems: namely, that if some particular past feature of a system is causally responsible for a present feature, an intermediary correlate must in principle be able to be individuated in the system. This assumption is false. The consequent model of deep learning systems, wherein explanations of behavior are intrinsically partial, yet complete and without remainder, has conceptual repercussions for various discussions surrounding explainable artificial intelligence.

Area 2 - Technology

Full Papers

Paper Nr:	47
Title:	Measuring Explainer Stability via Attribution Separability
Authors:	Eddie Conti, Álvaro Parafita and Axel Brando
Abstract:	Attribution methods (AMs) assign an importance score to each feature and are widely adopted to explain black-box models. However, most methods can produce variable attribution scores due to stochastic components in their definition. In this paper, we propose a distribution-based framework to capture the stability of attribution scores. In particular, our approach allows us to understand the degree of separability in the ranked attribution vector and obtain the largest index for which a feature ranking remains reliable. We further extend this framework to compare AMs based on the robustness of their rankings across a dataset. Through experiments, we demonstrate how to apply our method to evaluate explainer stability. Overall, our approach provides a complementary criterion for evaluating the stability of AMs.

Paper Nr:	54
Title:	Concept Bottleneck Models for Time Series with Applications to Human Activity Recognition
Authors:	Annemarie Jutte, Jelmer van der Graaf, Junseo Kim, Piotr Kondratowicz, Petru Manea, Milan Oosterink, Maria Secară, Harveer Singh and Faizan Ahmed
Abstract:	Predictive models for time series data are increasingly deployed across application domains. While deep learning models achieve high performance, their black box nature raises concerns about their reliability in high-stakes scenarios. Concept Bottleneck Models (CBMs) address these concerns by adding a layer of human-interpretable concepts. However, the application of CBMs to time series data remains limited, partly due to the challenges of defining meaningful concepts for (multivariate) signals. In this paper, we propose a methodology for developing CBMs for time series. Furthermore, we demonstrate how we can limit the model's reliance on spurious correlations by restricting concept predictions to physically relevant variables. Additionally, we propose validation methods for concept validity using data augmentation. We demonstrate our methodology on a use case from Human Activity Recognition using the MHEALTH dataset. The resulting CBMs achieve competitive predictive performance (85.26% accuracy with sensor masking, and 91.01% without), compared to a black box alternative (86.89%). While confirming whether spurious correlations are absent requires further evaluation, the results support CBMs as a potential alternative to black box models for safety-critical applications. The source code and results are available at https://github.com/SaxionAMI/CBM-for-time-series.

Paper Nr:	67
Title:	DICF: Domain-Invariant Counterfactual Explanations
Authors:	Ismail Bachchar, Françoise Fessant, Tassadit Bouadi and Thomas Guyet
Abstract:	One desired property of counterfactual explanations is plausibility-that is, their resemblance to real-world data. This property can be achieved by constraining counterfactuals to lie within the data manifold. Unfortunately, real-world data often come from multiple distinct environments (or domains) or change under distribution shifts. This may alter the counterfactual validity and plausibility. To address this issue, we propose Domain-Invariant Counterfactual Explanations (DICF), a novel self-explaining framework rooted in invariant representation learning. By explicitly disentangling the latent representation into an invariant class space and a domain space, DICF enables counterfactual search within a predictive space, while preserving the domain specificities. We also introduce a dedicated mechanism to learn the distribution of unseen domains. Empirical evaluations on several datasets demonstrate that when the underlying assumption of class and domain independence holds, such as in image data, DICF finds a trade-off between minimizing feature perturbations and maximizing domain preservation. Furthermore, in mixed-type tabular environments, where domain and predictive features are inherently entangled, the framework remains competitive against state-of-the-art baselines.

Paper Nr:	91
Title:	Interpretable Predictive Models Based on Statistical Element Networks Using Genetic Network Programming
Authors:	Shuhei Tanaami and Kaoru Shimada
Abstract:	In this study, we develop a novel XAI framework, termed the statistical element network, which constructs predictive models directly from interpretable and statistically grounded knowledge units. We refer to this method as the InterPretable Predictive Optimized Model based on Statistical Element Network (IPPOMSEN). This framework adopts a bottom-up approach in which the entire prediction process is built from human-understandable components, rather than by interpreting a pretrained global model. In this study, each statistical element is represented by an itemset with a statistically distinctive background (ItemSB), which explicitly encodes relationships between the combinations of attributes and the statistical distributions of continuous target variables. The set of ItemSBs is discovered through outcome-accumulation-type evolutionary computation. Genetic network programming is subsequently employed to optimize both the selection of ItemSBs and their transition structure. This enables the construction of an ItemSB network in which ItemSBs serve as statistical elements and support stepwise inference along a clearly traceable decision path. By restricting the number of statistical elements contributing to each prediction to at most three, the proposed method ensures strict interpretability while maintaining predictive performance. Experiments conducted on a geographical music dataset, which predicts latitude and longitude from timbral features, show that IPPOMSEN achieves the highest prediction accuracy among the compared methods, thereby outperforming conventional rule-averaging approaches and standard regression models. Experimental results indicate that statistical element networks provide an effective framework for constructing predictive models that balance interpretability and accuracy in applications that require transparent decision-making.

Paper Nr:	161
Title:	From Explainability to Interpretability: A Case Study in LLM-Based Behavioral Summarization Framework for Maritime Vessel Classification
Authors:	Alexander Tiu and Eric C. Larson
Abstract:	Interpreting the output of predictive models is a frequent and error-prone aspect of human–machine teaming. In maritime vessel classification, operators must understand not only why a model produced a prediction, but also what the supporting evidence implies about the vessel’s behavior. This work presents a modular large language model framework that transforms low-level model outputs into concise, operatororiented natural-language summaries. The proposed system combines multiple information modalities, including model classifications and confidence scores, SHAP-based feature analysis, and temporal and spatial feature aggregations, into a concise, glanceable summary for an operator investigating potential illegal fishing activity. These inputs are organized into a structured three-bullet summary describing the predicted vessel class, instantaneous behavior, accumulated behavioral history, and spatial trajectory information. To evaluate the contribution of each component, four ablated pipeline configurations were compared with the complete framework using a fixed cohort of 5,000 Automatic Identification System observations. Summary quality was assessed using a rubricbased language-model evaluator, with a separate blinded human evaluation used to examine agreement with the automated judgments. Under automated evaluation, the complete framework received the higher total rubric score in 65.4% to 86.0% of valid comparisons. The human evaluator also assigned the complete framework the highest mean score, although the two evaluators differed in their rankings of the individual ablations. These results indicate that transforming numerical model outputs into explicit behavioral descriptions before language generation produces more accurate, interpretable, and operationally useful summaries than directly presenting raw explainability information to a language model.

Paper Nr:	163
Title:	Topic-Aware Multi-Domain Sentiment Analysis with Explainable Counterfactual Interventions
Authors:	Vy Truong, Thu Le, Minh Huynh, Tung Le and Binh Dang
Abstract:	Sentiment expressions vary across domains, while many classification systems analyse text without explicitly representing topic context or providing sufficient insight into individual predictions. This paper introduces TESA-CF, a topic-aware and explainable framework for three-class sentiment analysis on a heterogeneous corpus of social-media posts and hotel-review sentences. A frozen Twitter-RoBERTa encoder produces 768-dimensional text representations through attention-maskaware mean pooling. An auxiliary TF-IDF–Logistic Regression classifier automatically predicts the topic of each input, which is mapped to a learnable 32-dimensional embedding and fused with the text representation in a neural classification head. LIME token attributions and an LLM-guided counterfactual module are further used to inspect selected predictions through targeted textual changes. TESA-CF is evaluated on a deduplicated corpus of 44,396 texts spanning 13 domains. The model achieves 87.93% accuracy and a macro F1-score of 0.8682, obtaining the best result among the models evaluated in this study, with a modest margin over the strongest TF-IDF baseline. Multi-seed ablation experiments show that BatchNormalization provides the largest performance contribution among the examined components, while the predicted-topic branch offers a moderate gain. In contrast, the skip connection produces no clear improvement under the current configuration. The LIME and counterfactual analyses provide complementary local evidence for examining model sensitivity, but are treated as instance-level diagnostics rather than proof of globally faithful reasoning.

Short Papers

Paper Nr:	59
Title:	Towards Trustworthy Human-Centered Manufacturing: A Decision Support Tool for {Model x Explainer} Selection
Authors:	Kaouther Benguessoum, Raoni Lourenço, Sylvain Kubler and Yves Le Traon
Abstract:	Explainable AI (XAI) is increasingly adopted in manufacturing to support human-in-the-loop decision-making, enabling practitioners to understand, validate, or override AI-driven recommendations. In practice, however, AI engineers and data scientists often select an explainer based on familiarity rather than evidence, overlooking the fact that the interplay among {dataset, model, explainer} substantially affects the trustworthiness of the generated explanation. This work presents a functionally grounded evaluation framework for multivariate time series (MTS) explanations, organized around three trustworthiness pillars: faithfulness, robustness, and complexity. We adapt state-of-the-art metrics from the image and tabular domains to time×channel attributions and integrate them into an interactive decision support system named TRUE-X (TRUstworthy EXplanations). Applied to an industrial hydraulic condition monitoring dataset across 20 {Model×Explainer} combinations, our results show that explainer choice has a substantial effect on trustworthiness scores, that no single explainer leads across all three pillars simultaneously, and that top-accuracy models do not consistently achieve Pareto-efficient trustworthiness configurations. TRUE-X enables practitioners to visualize and balance the trade-off between model accuracy and explanation trustworthiness, supporting more reliable and transparent XAI in human-centered manufacturing.

Paper Nr:	68
Title:	Region-Aware Fuzzy Surrogate Modeling for Explainable Glucose Prediction in Bioreactor Processes
Authors:	Arthur Babey, Christian Fischer, Sarah Lötscher, Karl Forner, Charlotte Baskerville, Gabriela Ivanova, Alessandro Di Cara and Carlos Peña
Abstract:	Global surrogate models explain a black-box model by training a simple, interpretable model to reproduce its predictions, and then reading the simple model as an explanation. Fidelity (how well the surrogate reproduces the black box) is usually reported as a single global number. This can hide the fact that the surrogate is accurate in some parts of the input space and misleading in others. We propose a region-aware fidelity analysis that splits the output range into regions where the surrogate can be trusted, where it can reveal why the black box fails, and where it should not be trusted. We apply this to glucose prediction in an industrial bioreactor dataset, using a FuzzyCoCo fuzzy inference system as a surrogate for a Random Forest. The surrogate reaches a test fidelity of R^2 = 0.91 and produces readable IF–THEN rules. Analyzing rule activations region by region shows that the model relies on a small set of competing rules driven by culture day and oxygen: it is accurate in the mid-range, where these two effects balance, and breaks down at the extremes, where one side dominates. Used as an independent check, SHAP points to the same two drivers, while the surrogate additionally shows how individual rules compete and how their balance shifts between regions.

Paper Nr:	69
Title:	ConvX: Efficient XAI for Real-Time Computer Vision with Single Forward-Pass Explanations
Authors:	Alex George, Lyudmila Mihaylova and Sean Anderson
Abstract:	In computer vision, many popular techniques for explainable AI (XAI) in multiclass classification are based on post-hoc analysis using computationally intensive backward passes through a network, e.g., Grad-CAM. In this paper, we introduce a fully convolutional method for explainable real-time image classification that operates with a high-speed single forward pass for all classes: ConvX. To accomplish this, we reformulate visual attribution as a multiclass multilabel hypersphere classification problem, embedding the attribution mechanism directly into the forward inference pathway of the network. In addition, our framework introduces an optional hybrid supervision mechanism (ConvX+) that incorporates dense pixel-level masks or sparse annotations (scribbles) into the loss function. Benchmarking results on the PASCAL VOC 2012 dataset show that our proposed method achieved a mean Intersection over Union (mIoU) of 0.35 using only image-level labels in a flat, deterministic runtime. ConvX achieved comparable localization performance to post-hoc alternatives such as Grad-CAM (0.35 mIoU), which took over twice as long to process a single class and slowed down linearly as the number of classes increased. Furthermore, using just five ground-truth masks per class, our hybrid variant achieved an mIoU of 0.48, substantially outperforming Grad-CAM and variants while maintaining the same low latency as ConvX. We also demonstrate that with five sparse scribble annotations per class, ConvX+ achieved an mIoU of 0.45, offering an attractive balance between annotation effort and accuracy. These results demonstrate that our framework is highly suited for deployment in edge environments such as robotics and industrial automation.

Paper Nr:	80
Title:	Right for the Wrong Reasons? Evaluating Plausibility and Behavioral Change in Explanation-Guided Learning
Authors:	David Dembinsky, Adriano Lucieri and Andreas Dengel
Abstract:	Explanation-Guided Learning (EGL) promises models that are not only accurate but also rely on task-relevant evidence. However, whether improved agreement between model explanations and human-provided rationales also implies a changed decision strategy remains unclear. We study this tension under controlled synthetic shortcut introduction using a constructed shortcut dataset. Inspired by real-world applications, our benchmark covers shortcuts of different kinds, including localized, background-based, and global cues. This allows us to evaluate whether different EGL objectives produce behavioral changes and explanatory changes. Across these axes, we find that the effects of EGL are strongly objective-specific. Rationale-based data augmentation provides the clearest robustness gains when shortcut evidence can be separated from task-relevant foreground information, reaching average accuracy improvements of up to +4.8 percentage points (pp) together with increased foreground reliance. Attribution-constraining objectives primarily improve explanation plausibility, increasing Importance Mass by up to +14 pp on average, but provide weaker and less consistent behavioral correction, with average accuracy gains of at most +1.6 pp. Moreover, neither improved robustness nor improved plausibility reliably implies higher perturbation-based fidelity. Overall, our results show that EGL can be useful, but its success cannot be inferred from rationale agreement alone and should instead be evaluated along complementary behavioral and explanationlevel axes. We publicly release all experimental results to enable further analysis (https://right-for-wrong-reasons.dfki.de/).

Paper Nr:	110
Title:	TRACE: Human-Centered Explanations for Audio Deepfake Detection via Perceptual Voice Traits
Authors:	Md Shajalal and Gunnar Stevens
Abstract:	Recent advances in neural speech synthesis and voice conversion have made audio deepfakes increasingly realistic, posing serious challenges for digital forensics and media authentication. While stateof-the-art anti-spoofing systems achieve impressive detection accuracy, they typically operate as opaque black boxes, providing little insight into the perceptual evidence underlying their decisions. We propose TRACE (Trait-Centric Explainable Detection), a human-centered framework for audio deepfake detection that explicitly aligns model explanations with perceptible voice traits. TRACE combines a high-performing Mel spectrogram based detector with a compact, interpretable trait representation capturing salient dimensions of human voice perception: tone and intonation, conversational flow and rhythm, energy dynamics, and spectral consistency cues. Detection performance is driven by spectral modeling, while explanations are generated exclusively in the trait space, ensuring attributions remain concept-level and human-interpretable by design. Experiments on the ASVspoof2019 LA benchmark demonstrate that the Mel+Traits configuration substantially outperforms classical GMM-based baselines. Ablation results confirm that perceptual traits function as an explanatory bottleneck rather than a standalone detector. Faithfulness is validated through deletion and insertion tests, showing that the identified traits are functionally relevant to the model’s predictions. Overall, TRACE demonstrates that high-performance audio deepfake detection can be paired with faithful, human-centered explanations grounded in perceptual voice concepts — addressing a critical requirement for forensic analysis and user-facing decision support.

Paper Nr:	127
Title:	ChartNarrator: Towards Faithful Chart-to-Text Generation for Financial Time Series
Authors:	Yutao Liu, Niki van Stein and Joost Visser
Abstract:	Faithful chart narration is an explainability task: each claim in the generated narrative must be traceable to a specific source constraint. Existing multimodal large language models (MLLMs) still struggle with numerical grounding, source-bounded reasoning, and reliable faithfulness evaluation in a financial time-series setting. We propose ChartNarrator, a framework combining structured anchor metadata, a two-stage generation pipeline, and a Kappa-verified LLM-as-a-Judge protocol. Analysis of 1,960 charts from CBS Open Data identifies four recurring faithfulness violations: Trend Direction Error, Numerical Faithfulness Violation, External Causal Hallucination, and Latent-state Mechanism Wording. Anchor metadata substantially improves generation quality, with the largest gain observed in Mechanistic Grounding. Human verification further shows that LLM-as-a-Judge reliability varies across evaluation dimensions, with the strongest agreement observed for Dependency Strength. In the final model performance comparison, the fine-tuned MLLM performs comparably to the Text-Only condition, while both outperform the No-Anchor and Zero-Shot baselines. These findings support the combined use of explicit source constraints, staged generation, and verified evaluation for faithful chart narration.

Paper Nr:	136
Title:	Explainable EMS Benchmarking for Renewable Energy Communities through Reproducible Multi-Microgrid Simulation
Authors:	Enrico De Santis, Gianluca Ferro and Antonello Rizzi
Abstract:	Renewable Energy Communities make energy management a multi-objective and multi-actor problem in which renewable generation, storage, internal exchange, tariffs and resilience interact across microgrids. Performance-only benchmarking is insufficient, because a controller may improve one indicator by depleting storage, increasing battery stress or exploiting assumptions hidden by aggregate scores. This paper proposes and positions an explainable and reproducible energy management system (EMS) benchmarking framework for Renewable Energy Communities based on controlled multi-microgrid simulation, explicit scenario semantics and storage-aware metric definitions. The framework evaluates six deterministic EMS strategies and three frozen supervised learned baselines under shared scenarios, price modes, forecast modes and seeds. The controlled benchmark shows that controller assessment changes across economic, grid-dependence, storage-use and resilience-oriented metrics: deterministic EMS rules remain competitive under several operating regimes, while learned frozen policies exhibit selective advantages that depend on the metric family used for evaluation. The resulting comparison characterizes controller behavior as metric-dependent, with rule-based, fuzzy-like, neural, tree-based and case-based controllers becoming interpretable when their information sets, training assumptions, traces and evaluation metrics are explicit. The developed platform supports advanced teaching by turning controller rationale, learned-policy auditability and energy-management trade-offs into inspectable objects.

Paper Nr:	156
Title:	Towards Safety-Aware Edge AI for Parkinson's Tremor: A Feasibility Study of Interpretable and Learned Models with Out-of-Distribution Monitoring
Authors:	Ilya Afanasyev
Abstract:	On-device tremor quantification could support closed-loop deep brain stimulation for Parkinson’s disease, but it must balance discrimination, robustness, latency, and safety. Reported performance for wearable tremor detection often rests on randomized splits that leak patient-specific information. We benchmark nine feature-based models and a raw-signal 1D-CNN on TremorDB (laser-velocimetry finger kinematics, 13 two-class subjects) under strict Leave-One-Subject-Out (LOSO) validation. LOSO raw accuracy is about 72%, but this figure overstates discrimination: at the default threshold no model exceeds a balanced accuracy of 0.53, and five models, the 1D-CNN among them, are pure majority-class predictors (balanced accuracy 0.500). ROC-AUC is uniformly weak (0.57 to 0.63), and no pairwise ROC-AUC or F1 difference survives multiple-comparison correction (Wilcoxon, Holm-corrected p at least 0.55). Class re-weighting raises balanced accuracy only on four high-amplitude subjects and leaves ROC-AUC unchanged. A subject-blind random split in the same pipeline inflates ROC-AUC to 0.76 and balanced accuracy to 0.60 but never approaches the accuracies exceeding 99% reported in the literature, so leakage alone cannot explain them. We use the Explainable Boosting Machine to diagnose the failure: its additive structure shows attention to amplitude dispersion (MAD) and 4 to 6 Hz tremor-band features yet no generalization, identifying inter-subject variability as the bottleneck. Unsupervised per-subject normalization does not recover discrimination (0.63 to 0.55), consistent with a higher-order shift that calls for supervised few-shot personalization. An Isolation Forest monitor gives failsafe gating but is validated only on synthetic noise and fails on 2 of 15 subjects. We report this as a negative-result feasibility study: without personalized calibration, patient-agnostic tremor classifiers default to majority-class guessing, and transparency makes this failure diagnosable.

Paper Nr:	36
Title:	Comparing Traditional and AutoML Approaches: Performance Evaluation and Explainability Insights
Authors:	Alessio Carachino, Davide Storelli, Enza Giangreco and Marco Pinnella
Abstract:	Explainability is a critical requirement in sensitive domains such as the public sector, where transparency and trust underpin effective decision-making. In public health, for instance, predictive models can significantly enhance outcomes but often rely on traditional machine learning (ML) pipelines that demand extensive human intervention, making them resource-intensive and susceptible to bias. Automated machine learning (AutoML) frameworks, such as AutoGluon and H2O, address these limitations by automating key stages of the ML workflow, thereby improving efficiency without compromising performance. This study compares traditional ML approaches with AutoML solutions for heart disease prediction, evaluating both predictive performance and interpretability. Model explainability is assessed using SHAP, a widely adopted model-agnostic Explainable AI technique that provides insights into feature contributions driving the predictions. Results demonstrate that AutoML-based methods effectively manage complex predictive tasks while maintaining transparency. The feature importance derived from SHAP closely aligns with clinical risk factors reported in the literature, reinforcing the reliability of AutoML as a decision-support tool. These findings suggest that AutoML can serve as a robust and interpretable solution for data-driven decision-making in sensitive and high-impact domains where transparency and reliability are essential.

Paper Nr:	48
Title:	A Method for Explainable Automated Patent Classification by Means of Transformer Models
Authors:	Sara Poce, Francesco Mercaldo, Giulia Varriano, Fabio Martinelli and Antonella Santone
Abstract:	Patent documents represent a primary source of knowledge for professionals belonging to both corporate and research environments, and their classification according to taxonomies, such as Cooperative Patent Classification (CPC), is fundamental for Intellectual Property-related tasks such as prior art search, technology monitoring and forecasting. However, with the consistently growing number of patent filings, together with the challenges characterizing this cumbersome task, traditionally performed by patent examiners, makes the manual classification an overwhelming practice, motivating the need for automated approaches. We propose a Transformer-based patent CPC section-level classification method in order to automatize this task. Experimental results achieve an Accuracy and F1-score of 0.832 and 0.831, respectively. Moreover, due to the black-box nature of these models, Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP) explainability techniques are considered with the aim of providing a transparent and trustworthy classification, by demonstrating that the proposed method leverages domain-specific terms aligned with the target CPC section, while lexical ambiguities across different CPC sections represent the primary source of misclassification.

Area 3 - Applications

Full Papers

Paper Nr:	73
Title:	Explainable AI for SME Financial Distress Prediction: From Model Comparison to Supplier-Risk Decision Support in Shipbuilding Networks
Authors:	Inoka Ranatunga, Pertti Ranttila and Kimmo Tarkkanen
Abstract:	Small and medium-sized enterprises (SMEs) play a critical role in industrial supply networks, yet their financial distress may cause disruptions that extend far beyond the focal firm. In long-cycle shipbuilding projects, financially vulnerable suppliers can contribute to delivery delays, re-sourcing needs, material-management inefficiencies, and reduced network resilience. This study develops and compares four predictive models for SME financial distress prediction - Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a Deep Neural Network (DNN) - using the Polish Companies Bankruptcy dataset. To improve robustness, the modelling pipeline applies median imputation, outlier treatment using the 3×IQR rule, z-score normalization, class weighting, and SMOTE. XGBoost achieved the strongest overall predictive performance on unseen data. To address the black-box problem and improve practical usability, SHAP-based explainable AI (XAI) was applied to interpret the best-performing model at both global and local levels. The results show that profitability, liquidity, and cost-efficiency indicators are the most influential drivers of financial distress, with operating-profit coverage of financial expenses emerging as the strongest predictor in this dataset. Beyond model comparison, the paper positions explainable financial prediction as a transparent decision-support component for supplier-risk management in shipbuilding networks. SHAP-based explanations make it possible to identify whether elevated supplier risk is primarily associated with weak liquidity, insufficient operating-profit coverage, or cost pressure, thereby supporting more transparent supplier selection, ongoing monitoring, and resilience-oriented risk management. The study contributes to XAI research by demonstrating how explainable financial models can be connected to broader sustainability and operational decision-support contexts.

Short Papers

Paper Nr:	30
Title:	Natural Language Explanations in News Rankings: Evaluating Large Language Model-Generated Explanations in a Borda Count News System
Authors:	Mario Villar, Thorsten Zylowski, Matthias Wölfel, Noelia Rico and Irene Díaz
Abstract:	Decision-making tools often rely on ranking aggregation methods, such as the Borda count, to combine multiple criteria into an aggregated outcome. However, these systems frequently lack explanations for non-expert users. Recent advances in large language models (LLMs) enable the generation of natural language explanations that can improve transparency and user understanding, particularly in digital news contexts where misinformation is prevalent. This paper investigates explanations in a news ranking system through a user study comparing a scenario without explanations, one with numerical explanations, and one with LLM-generated explanations. Additionally, we present a series of LLM prompts to generate natural-language explanations in systems based on the Borda count across multiple contexts, and introduce a scraping tool to collect up-to-date, fact-checked data from the PolitiFact website. The results show that LLM-generated explanations improve perceived understanding, decision support, and awareness of misinformation, although they did not have an influence on reducing fake news consumption. In addition, users expressed a desire for receiving natural language explanations in newspapers, highlighting their potential to enhance transparency in such systems.

Paper Nr:	35
Title:	Interpretability-Based Hallucination Detection for RAG-Based LLMs in Cybersecurity
Authors:	Clément François and Anne-Laure Wozniak
Abstract:	Hallucination detection in large language models (LLMs) has become an important area of research. In this paper, we propose a practical hybrid framework (Code available at: https://github.com/Kereval35/Hallucination-Detection-Framework) for improving the reliability of the outputs in retrieval-augmented generation (RAG)-based LLM systems, based on methods derived from both interpretable internal model representations and external evaluation mechanisms. Our approach combines internal signal analysis through probe classifiers, with external evaluation by an ML-as-a-judge hallucinations detector. It is validated in a representative cybersecurity use case, demonstrating its applicability in sensitive, domain-specific settings at inference time. Our findings indicate that both RAG configurations and prompt formulation significantly influence LLMs performance, indicating that prompting strategies can influence not only the model behaviour but also the internal representations associated with the detection of hallucinations. At the same time, the results of the probes reveal that internal model activations encode signals indicative of contextual irrelevance, underscoring the value of interpretability-driven methods for hallucination detection.

Paper Nr:	106
Title:	Quantifying Feature Importance Drift in Machine Learning Regression Models Induced by Missingness and Imputation
Authors:	Satyam Dudhagara, Lukas Hunold, Parsha Pahlevannejad, Achim Wagner and Martin Ruskowski
Abstract:	Missing data commonly affect industrial machine-learning pipelines combining sensors, simulations, process logs, and laboratory measurements. Although imputation enables model training, it can alter the learned statistical structure and affect post-hoc explanations. Prior work has examined predictive performance and explainability, but feature-importance drift remains understudied in industrial regression with process-dependent missingness. In this paper, we investigate how missingness and missing-data handling strategies influence Shapley Additive Explanations (SHAP) -based global feature importance in an XGBoost regression model for yield-strength prediction from industrial forging data. We compare unstructured Missing Completely at Random (MCAR) masking with process informed feature-subset missingness concentrated in process- and material-related variables. Four missing-data handling strategies are evaluated: median imputation, k-nearest neighbor (KNN) imputation, multivariate imputation by chained equations (MICE), and XGBoost’s native missing-value handling. The 24 conditions were each repeated with 10 matched random seeds, yielding 240 experiments. Predictive performance and attribution stability were evaluated per repetition and reported as mean and sample standard deviation. Attribution stability was assessed using Spearman rank correlation, top-five Jaccard overlap, and relative ℓ1 SHAP drift. The results show that KNN generally provides the strongest combination of predictive performance and attribution stability. However, under process-informed feature-subset missingness at 50%, median imputation achieves the lowest mean RMSE, whereas KNN preserves the feature ranking most effectively. Across all 24 condition-level means, RMSE and SHAP rank stability were only weakly associated. These findings demonstrate that predictive performance alone does not establish the stability of model explanations under missing data.

Paper Nr:	124
Title:	“I’ll Be Crystal Clear with You!”: xAI-Equipped Conversational Agent for Smart Manufacturing
Authors:	Alexandra Kovacs, Stefano Carrino, Corinne Hager Jörin, Bernard Ries and Elena Mugellini
Abstract:	In smart manufacturing, predicting machine failures or machining quality degradation is valuable, but insufficient without identifying root causes. Artificial intelligence (AI) can support Industry 4.0 by assisting human workers. However, for AI-driven automation to be accepted, explainable AI (xAI) is required to ensure transparency, provide human-understandable explanations, and earn user trust. Since explanations are dialogic, we developed a conversational agent (CA). Currently, xAI in conversational systems for industrial contexts remains limited. We present our journey toward developing an xAI-equipped CA for machine operators, building on prior research on the xAI question bank (xQB): we move xQB from concept to practice in manufacturing, provide an integration blueprint and evaluate it on a dataset collected from a production machine. We employ and integrate xAI techniques and large language models (LLMs) into a smart manufacturing CA to generate explanations. We share design takeaways and show that conversational explanations can link predictions to root causes.

Paper Nr:	166
Title:	Localization Gains from Explanation Regularization Should Not Be Interpreted as Faithfulness Gains
Authors:	Amrin Akter, Mauro Dalla Mura, Diana Mateus and Dawood Al Chanti
Abstract:	Training-time explanation regularization is often evaluated by whether saliency maps become more concentrated within annotated regions. We argue that this criterion alone is insufficient: transformation consistency, lesion localization, and faithfulness to the classifier are distinct properties and should be evaluated separately, particularly in medical imaging. We examine this distinction through a case study of Contrastive Grad-CAM Consistency (CGC) in three-class brain tumor MRI classification. Using a single-channel ResNet-50 with five-fold patientgrouped cross-validation, we compare a cross-entropy baseline with CGC. CGC changes overall Pointing Game (PG) from 27.24% to 36.02% and Relevance Mass Accuracy (RMA) from 11.17% to 18.01%, while classification accuracy changes from 93.63% to 94.87%. Perturbation-based faithfulness metrics show inconsistent changes: Insertion Area Under the Curve (IAUC) changes from 0.73 to 0.78, whereas Deletion Area Under the Curve (DAUC) favors the baseline (0.36 versus 0.41; lower is better), and overall Iterative Removal Of Features (IROF) is 0.57 for both models. These results support our position that improved spatial agreement with pathology should not be interpreted as a general improvement in explanation faithfulness. We therefore recommend property-specific evaluation in which the property targeted by the explanation objective, spatial agreement with reference regions, and model reliance are measured and reported separately.

Paper Nr:	33
Title:	Classical Planning Meets Large Language Models: An Automated Evaluation Framework for Workflow Plans
Authors:	Chanuka Inupa Bandara and Prasan Yapa
Abstract:	Large Language Models (LLMs) have recently demonstrated impressive capabilities in generating multi-step workflows and plans from natural language descriptions. However, evaluating the quality of such plans remains a fundamental challenge. Existing evaluation approaches often rely on surface plausibility, or subjective human judgment, which fail to assess whether generated plans are executable or how close they are to optimal solutions. In this study, we propose an evaluation framework for evaluating the quality of multi-step workflow plans generated by LLMs. We observe that classical planning systems provide strong correctness and optimality guarantees, making them well-suited as objective evaluators for LLM-generated plans. We introduce a pipeline in which LLMs translate natural language task descriptions into formal Planning Domain Definition Language (PDDL) problem specifications. Classical planners are then used to generate executable plans, which are formally validated and compared against optimal solutions. Plan quality is quantified using two complementary metrics: plan validation rate, which measures executability, and optimality gap, which captures deviation from the optimal solution. We conduct extensive experiments across five standard planning domains using three open-source LLMs under varying decoding temperatures. Our results reveal a consistent trade-off between decoding temperature and plan quality, with lower temperatures yielding higher executability across models and domains. We further observe strong domain and model dependent behavior, where different LLMs succeed in different planning domains. These findings demonstrate that executability and optimality must be jointly evaluated to reliably assess LLM-generated workflow plans.