CoopIS 2026 Abstracts

Area 1 - Applications of AI-Augmented and Autonomous Information Systems

Full Papers

Paper Nr:	71
Title:	Quantifying Clinical Risk in Generative Radiology Reporting: MCRS, a Medical Composite Reliability Score
Authors:	Wala Tlili, Samawel Jaballi, Mounir Zrigui and Henri Nicolas
Abstract:	Currently, the evaluation of AI-generated radiology reports is based on surface-level metrics (BLEU, ROUGE, and METEOR) that measure token overlap with no means of clinical fact checking. In a con- trolled five-system benchmark on the Indiana University Open-I corpus (n = 500 test instances per system), we show that the metrics are clin- ically unreliable: the Spearman correlation between ROUGE-L and our proposed metric is only ρ = 0.316 (p = 1.5 × 10−47, n = 2,000). The two rankings share less than 10% of their rank variance, with ρ2 ≈ 0.10. To fill this gap, we propose MCRS (Medical Composite Re- liability Score), a composite score ∈ [0, 1] that considers four comple- mentary clinical dimensions: Claim-Fact Consistency (CFC, 30%), Clinical Grounding Score (CGS, 25%), Reasoning Consistency Score (RCS, 20%), and a Risk-Weighted Hallucination Score (RWHS, 25%). The weights were elicited from clinical experts, and the contribution of each component was confirmed post-hoc by ablation. Our benchmark reveals a systematic rank inversion:Qwen2.5-7B ranks 2nd by ROUGE-L among the four non-fine-tuned systems yet last by MCRS, with critical-severity hallucinations in 52.8% .This danger is totally un- seen by surface metrics. For this purpose, we propose an adapted four- class hallucination taxonomy for radiological risk. Wilcoxon signed-rank tests confirm all pairwise differences (p < 10−55, r > 0.40, large effect). We present MCRS as a clinical and safety focused evaluation metric and demonstrate its correlation with expert radiologist ratings

Paper Nr:	117
Title:	CalendarBench: Benchmark Framework for Personal Scheduling in the Wild
Authors:	Maani Beigy, Roy van den Munckhof, Laura Genga and Pieter Van Gorp
Abstract:	Effective planning and managing personal time are associated with higher productivity and greater well-being. Improvements in health, happiness, and personal development depend on whether people allocate time to beneficial activities, which often requires adopting new routines that challenge existing habits. Planning when to perform each new activity and assessing their cognitive and physical demand can be a burden. This challenge has motivated an increasing number of tools and techniques that automate personal scheduling, from constraint solvers to planners and intelligent personal assistants. However, existing benchmarks do not evaluate single-user, multi-objective personal task scheduling with dynamically generated tasks and soft, human-centric metrics. This limits controlled comparisons of scheduling methods in this knowledge-intensive setting before studies with human participants. We address this gap with CalendarBench, a configurable benchmark framework that provides knowledge-grounded, persona-based simulation of a user’s calendar and context trace. We use large language model (LLM) capabilities and graph retrieval augmented generation (GraphRAG) to enable dynamically generating recommended activities and evaluating any scheduled plan with a multi-objective human-centric metric. A scheduler of interest can be compared with state-of-the-art (SOTA) techniques using built- in configurable comparisons. We demonstrate such capabilities with an example benchmark experiment and testing with five SOTA models.

Paper Nr:	136
Title:	A Multi-Agent Simulation and Graph-Based Deep Learning Approach for Dynamic Resource Coordination in Forest Fire Crises
Authors:	Hiba Jouini, Ines Thabet and Dhouha Maatar
Abstract:	The increasing frequency and intensity of forest fires, particularly in Mediterranean regions such as Tunisia, reveal the limitations of traditional crisis management systems, particularly in terms of realtime coordination and resource allocation. An effective response requires the rapid and consistent allocation of heterogeneous emergency resources within operational constraints. This article proposes RoboGNN, a decision support approach based on multi-agent systems (MAS) and graph neural networks (GNN) for managing forest fire crises. The crisis environment is modeled as a graph representing emergency units, civilians, and fire fronts with their spatial and functional relationships. A graph attention network (GAT) is trained by imitation learning from an optimization-based expert solver and evaluated in a NetLogo-based simulation environment on crisis scenarios inspired by the Tunisian context. Experimental results show that RoboGNN outperforms three baselines: the greedy heuristic, Min-Cost Flow (MCF) optimization, and Multi-Layer Perceptron (MLP)-based imitation learning, highlighting the benefits of graph-based modeling. Compared to the greedy heuristic, RoboGNN achieves a reduction of 23.5% in intervention time, a decrease of 30.0% in burned areas, and a 66.7% reduction in civilian casualties. Compared to MCF, RoboGNN reduces intervention time by 8.9% and burned areas by 9.6%, while achieving comparable civilian safety outcomes. Compared to the MLP baseline, RoboGNN reduces intervention time by 10.6%, burned areas by 13.1%, and civilian casualties by 19.0%, demonstrating that the performance gains stem from the relational reasoning introduced by the graph structure.

Short Papers

Paper Nr:	35
Title:	GTAFNet: An AI-Augmented Architecture for Low-Latency Assistive Healthcare Communication
Authors:	Sirine Ammar, Nesrine Triki, Mohamed Karray and Mohamed Ksantini
Abstract:	Brain-Computer Interfaces (BCIs) are vital assistive information systems for individuals with severe motor impairments. However, P300-based spelling BCIs suffer from weak signal-to-noise ratios and require numerous stimulus repetitions, severely reducing communication throughput. This paper proposes GTAFNet, a human-centered, AI-augmented architecture to enhance P300 communication speed and reliability. Instead of purely data-driven representations, GTAFNet incorporates structured biomedical knowledge by formalizing the 10-20 electrode system as a fixed anatomical spatial prior graph. A Graph Attention Network (GAT) exploits this prior to propagate information via dynamic, trial-specific attention weights along neurophysiological pathways. To our knowledge, this is the first architecture combining xDAWN spatial filtering with anatomically informed GATs. Evaluated on eight subjects (31,200 trials each), GTAFNet achieves a mean character accuracy of 0.974 at 15 repetitions, with high accuracy reliably reached using only 3–5 repetitions, whereas benchmarks (LDA, ESVM, EoCNN) require 15 repetitions for 85–90% accuracy. Mean AUC is 0.802 with an information transfer rate of 33.2 bits/min at one repetition. By drastically reducing decoding time, GTAFNet advances practical, low-latency AI-augmented healthcare systems.

Paper Nr:	92
Title:	Beyond Heuristics and Commercial Navigation: An AI-Augmented Decision-Support System for Ambulance Travel Time Estimation
Authors:	Muhammad Adil Abid, Johan Holmgren, Fabian Lorig and Jesper Petersson
Abstract:	Accurate ambulance travel time estimation is critical for emergency medical dispatch, yet existing decision-support practices often rely on static lookup tables or simplified assumptions regarding how ambulance travel time differs from ordinary passenger-car travel. For instance, they assume ambulances simply travel a fixed percentage (e.g., 5%) faster than ordinary passenger cars. However, such assumptions are rarely validated using real ambulance operational data and may fail to capture the effects of trip priority, temporal context, and origin–destination. We address these limitations using historical ambulance trips from the Skåne region in Sweden. Our analysis reveals that ambulance travel times vary substantially by hour, day, and trip priority, and that commercial navigation systems fail to capture emergency-specific dynamics. We contribute a machine learning system to more accurately estimate ambulance travel time. To make the system interpretable, we explored SHapley Additive exPlanations and permutation importance techniques to ensure that our system captures operationally meaningful patterns.

Area 2 - Architecture, Engineering, and Management of AI-Enabled Information Systems

Full Papers

Paper Nr:	39
Title:	Data Product Provenance: Enabling Trustworthiness in Cooperative Information Systems
Authors:	Laura Schuiki, Tanja Auge, Christoph Stach, Corinna Giebler, Eva Hoos, Meike Klettke and Bernhard Mitschang
Abstract:	The increasing adoption of the data product approach in cooperative information systems is transforming how data is handled. Unlike traditional datasets, data products encapsulate data with code, infrastructure, and metadata, enabling interoperable data-driven services in decentralized ecosystems. While this paradigm fosters innovation, it also introduces new challenges for transparency, accountability, and regulatory compliance. For traditional data handling, provenance can be used, however, existing approaches are either too generic to capture specific characteristics of data products or too narrowly tailored to individual use cases, limiting their applicability in cooperative environments. This paper introduces Data Product Provenance to enable trustworthy data products. Building on existing provenance concepts, we investigate how these concepts can be transferred to the context of data products in order to represent the complex multi-stakeholder interactions present in cooperative information systems. To this end, we introduce an areaagnostic and use-case-independent, yet data-product-specific ontology, called DPPROV, which captures the distinct lifecycle phases and characteristic properties of data products, resulting in a concept for data product provenance. We assess DPPROV based on design requirements as well as insights from experts at a globally operating manufacturer. Their feedback indicates that DPPROV provides a practical foundation that strengthens transparency, regulatory compliance, and trust in cooperative information systems.

Paper Nr:	41
Title:	TRACE: Hybrid Retrieval and Deterministic Tool Orchestration for Auditable AI Compliance Systems
Authors:	Amerigo Konrad Velletti
Abstract:	Multinational enterprises face growing pressure to maintain transfer pricing compliance as tax authorities digitize enforcement and regulations become more complex. Generative artificial intelligence offers the potential to automate documentation tasks, but standard large language models (LLMs) struggle with hallucinations and often fail to strictly follow legal rules. This paper addresses these limitations by designing and implementing an agentic Retrieval-Augmented Generation (RAG) system, TRACE, that automates intercompany loan compliance under German transfer pricing law (AStG and VWG VP). The system combines structured financial data with a semantic index of regulatory texts across a six-layer architecture. A ReAct agent equipped with domain-specific tools orchestrates the pipeline, handling naturallanguage queries, structured database lookups, semantic retrieval, and deterministic compliance calculations. Evaluation through a controlled ablation study shows that a hybrid retrieval strategy reaches 91.7 % context coverage, outperforming rule-based routing by 41.7 percentage points. On a benchmark of 25 standardized scenarios, TRACE achieves 100 % accuracy for quantitative compliance decisions. However, an LLMas-Judge evaluation indicates an average output quality score of 2.51 out of 5, exposing a significant gap between high retrieval coverage and explanation quality. The results demonstrate the viability of agentic RAG for automating tax compliance while identifying key areas for future improvement.

Short Papers

Paper Nr:	37
Title:	Enhancing Federated Learning Robustness through Hierarchical Edge-Fog-Cloud Architecture and Swarm-Evolutionary Optimization
Authors:	Wassim Rahmoun, Samir Ouchani, Nassima Dif and Abdelhak-Djamel Seriai
Abstract:	Randomised Optimisation Algorithms (ROAs) are widely used to solve complex optimisation problems due to their simple design and implementation and improved average-case performance. However, their effectiveness depends strongly on the design and interaction of their algorithmic components. Properly selecting algorithmic components is a very effortful task due to the increasing number of possible components, such as operators and mechanisms, found in the literature. Therefore, evaluating various combinations of components is essential to reveal their effects on the algorithm’s behaviour and performance. This study focuses on an established ROA, i.e., the Genetic Algorithm (GA), and investigates the behaviour of its components on a combinatorial problem, i.e., the capacitated Vehicle Routing Problem (VRP). The work considers alternative mechanisms for solution representation, route construction, parent selection, recombination, and mutation. A set of benchmark instances with different customer-distribution characteristics is used to evaluate the effect of these components on solution quality, convergence behaviour, and population diversity. The experimental results show that some component choices have a more consistent influence on performance, while others exhibit more instance-dependent behaviour. Overall, the study highlights the importance of analysing GA components not only individually, but also in terms of their compatibility with the solution representation and problem structure. The findings provide useful insight for the design of modular and reusable GA-based optimisation frameworks.

Paper Nr:	133
Title:	Towards Unified Data and AI Governance: A Systematic Literature Review
Authors:	Erik Petsch and Manuel Fritz
Abstract:	The EU AI Act and related frameworks require organizations to maintain governance evidence throughout the AI lifecycle. In AI-enabled cooperative information systems, where data products, models and decision services are owned by different teams and exchanged across trust boundaries at run time, that evidence is produced by one actor and needed by another, yet research on it remains fragmented. We conduct a concept-centric systematic literature review for 2015--2026, synthesizing 49 publications of which 44 are coded individually in a Webster--Watson concept matrix. The resulting taxonomy of seven governance capability areas shows that the literature offers building blocks but no end-to-end approach linking architecture-spanning data/model interfaces with lifecycle triggers, audit evidence and machine-readable regulatory rules. Only five coded sources address decentralized or hybrid settings, and only three treat data and models as continous governance objects. From this gap we derive three desiderata for governance infrastructures that aim to keep evidence traceable throughout the complete lifecycle.

Paper Nr:	133
Title:	Towards Unified Data and AI Governance: A Systematic Literature Review
Authors:	Erik Petsch and Manuel Fritz
Abstract:	The EU AI Act and related frameworks require organizations to maintain governance evidence throughout the AI lifecycle. In AI-enabled cooperative information systems, where data products, models and decision services are owned by different teams and exchanged across trust boundaries at run time, that evidence is produced by one actor and needed by another, yet research on it remains fragmented. We conduct a concept-centric systematic literature review for 2015--2026, synthesizing 49 publications of which 44 are coded individually in a Webster--Watson concept matrix. The resulting taxonomy of seven governance capability areas shows that the literature offers building blocks but no end-to-end approach linking architecture-spanning data/model interfaces with lifecycle triggers, audit evidence and machine-readable regulatory rules. Only five coded sources address decentralized or hybrid settings, and only three treat data and models as continous governance objects. From this gap we derive three desiderata for governance infrastructures that aim to keep evidence traceable throughout the complete lifecycle.

Paper Nr:	134
Title:	Designing Event-Centred AI Visual Monitoring for Smart Manufacturing: A Design Science Study in Wood-Panel Production
Authors:	Francisco Rodrigues, José Monteiro, Luís Veloso, Leonardo Porto and João Barata
Abstract:	Manufacturing information systems need to turn continuous camera streams into records that supervisors can inspect and interpret. This Design Science Research deploys an event-centred visual monitoring artefact in wood-panel production. The design combines contextual object detection, task-specific temporal rules, persistent metadata, image evidence and a web dashboard to monitor material handling and inspect stacked panel batches. Internal validation yielded precision/recall of 0.962/0.999 for vehicle detection and 0.763/0.509 for anomaly detection, on 45 and 92 images, respectively. Functional demonstrations established the integration of detection, event registration and evidence retrieval. The study distinguishes interval-based operation records from anomaly observations and presents three provisional design principles grounded in documented requirements and implementation experience. An affordance perspective connects these mechanisms to potential supervisory actions. The evidence supports technical feasibility in the observed setting; independent event-level and user evaluations provide the next steps towards assessing operational effectiveness.

Area 3 - Ethical, Privacy-Preserving, and Secure Information Systems

Full Papers

Paper Nr:	111
Title:	LANTERN: A Cost-Aware Pipeline for Web Attack Detection with LLM Verification and Semantic Caching
Authors:	Truong-An Vu, Juba Agoun, Clara Bertolissi, Laurent d'Orazio, Sébastien Hervieu and Thi-Huong-Giang Vu
Abstract:	High-volume HTTP traffic makes web attack detection a challenging trade-off between detection quality, computational cost, and analyst usability. While rule-based and machine learning detectors provide efficient first-line filtering, they may miss sophisticated attacks or generate difficult-to-interpret alerts. Large Language Models (LLMs) can improve verification and produce natural-language explanations, but their computational cost prevents their deployment on every request. This paper presents LANTERN, a cost-aware architecture for scalable HTTP attack detection. Rather than replacing conventional detectors, LANTERN reserves LLM reasoning for requests requiring deeper analysis, reusing decisions for semantically similar inputs. We formulate threshold selection as a cost optimization problem that jointly controls alert escalation and cache reuse. We evaluate LANTERN on the public WebAttack-CVSSMetrics and CSIC 2010 datasets under both in-distribution and distribution-shift settings. The results show that LANTERN maintains high detection performance while substantially reducing the number of LLM inferences through semantic caching, providing an effective trade-off between detection quality, explanation capability, and operational cost.

Short Papers

Paper Nr:	13
Title:	AP-FedFCA: Adaptive Privacy-Aware Federated Formal Concept Analysis
Authors:	Saida Sfaxi, Mokhtar Sellami and Mohamed Mohsen Gammoudi
Abstract:	Federated Formal Concept Analysis (FedFCA) enables collaborative knowledge discovery from distributed formal contexts without centralizing raw data. However, a uniform privacy budget may provide an inadequate privacy–utility trade-off when local contexts differ in structure and distribution. We propose Adaptive Privacy-Aware Federated FCA (AP-FedFCA), which dynamically assigns a local privacy parameter to each participant according to context density and structural attribute variability. Each local incidence matrix is privatized using randomized response before constructing the corresponding concept lattice. The resulting local lattices are then aggregated through closurepreserving operations. AP-FedFCA is evaluated on Mushroom, PUMSB, Retail, and Kosarak under both IID and Non-IID partitions using Global Stability, Concept Preservation Rate, Support Degradation, and Average Jaccard Similarity. Experimental results show that AP-FedFCA generally provides better structural preservation than fixed-budget FedFCA and FedDP-FPM, particularly under heterogeneous data distributions.

Paper Nr:	28
Title:	Identifying GDPR-Critical Tasks in Business Processes Using Large Language Models
Authors:	Magdalena von Schwerin, Merten Dieckmann and Manfred Reichert
Abstract:	Ensuring GDPR compliance in business processes is resource-intensive, as manual oversight is error-prone and costly. We investigate whether LLMs can reliably identify GDPR-critical activities in BPMN 2.0 process models in a zero-shot setting and present GRIPL (GDPR Risk Identification in Processes using LLMs), evaluated on a newly constructed dataset of 93 real-world BPMN models with activity-level GDPR annotations across six domains. A systematic benchmark of ten open-weight and proprietary LLMs shows that European open-weight models (≥12B parameters) consistently achieve F1 ≥0.80, a viable alternative to proprietary APIs for data-sovereign deployment. However, while several LLMs achieve high recall (>0.90), precision drives performance differences across domains and languages. These results show that LLM-based classification can reliably flag GDPR-critical activities, providing a promising basis for GDPR pre-screening in legal workflows.

Paper Nr:	54
Title:	Securing Personal Visual Identity: A Graph Attention Network Approach for Multi-Face Deepfake Detection
Authors:	Hanadi Elsablaoui, Mohamed Amine Mezghich, Ridha Ghayoula and Lassaad Latrach
Abstract:	The integrity of personal information in digital systems is increasingly threatened by deepfake technology, which enables sophisticated visual identity forgery and large-scale misinformation. Current deepfake detection approaches treat each face independently, failing to exploit relational context among multiple faces, a critical limitation in real-world information systems where visual content involves multiple individuals. We propose a graph-based authenticity verification framework that models inter-face dependencies to protect personal visual identity in information systems. Each face, detected by YOLOv8, is represented as a graph node with CNN-based feature embeddings. Graph Neural Networks including Graph Attention Networks (GAT) and GraphSAGE perform collective relational reasoning to classify facial content as real or manipulated. Extensive experiments show that GAT combined with EfficientNet-B0 achieves 88.23% accuracy and 87.21% F1-score, outperforming single-face baselines. Cross-dataset evaluation further confirms the generalization capability of the proposed approach, establishing it as a reliable component for privacy-preserving and trustworthy visual information systems.

Paper Nr:	62
Title:	Byzantine-Resilient Federated Formal Concept Analysis
Authors:	Saida Sfaxi, Mokhtar Sellami and Mohamed Mohsen Gammoudi
Abstract:	Formal Concept Analysis (FCA) provides an interpretable framework for deriving formal concepts and concept lattices from binary relations. Federated Formal Concept Analysis (FedFCA) extends FCA to distributed environments without sharing raw local data, but it remains vulnerable to Byzantine participants that may inject fake concepts, manipulate support or confidence values, submit structurally inconsistent updates, or collude. We propose BR-FedFCA, a Byzantine-resilient extension of FedFCA based on a server-side validation layer combining FCA structural verification, support and confidence consistency analysis, Byzantine scoring, and trust-aware filtering. Experiments on Mushroom, Chess, and T10I4D100K under IID and skewed Non-IID distributions evaluate runtime scalability, Byzantine detection, attack success rate, and structural robustness. Results show improved robustness under the considered synthetic attacks while maintaining practical scalability for offline or batch-oriented federated knowledge discovery. The implementation is publicly available on GitHub (https://github.com/saida-coder/Byzantine-Resilient-Federated-Formal-Concept-Analysis).

Paper Nr:	116
Title:	Decentralized Identity in Healthcare: A Systematic Security and Privacy Assessment
Authors:	Ruben Mbiyavanga, Walid Fdhila, Nizar Messai and Yacine Sam
Abstract:	Healthcare is increasingly cited as a prime application domain for Self-Sovereign Identity (SSI), where Decentralized Identifiers (DIDs), Verifiable Credentials (VCs), and distributed ledgers promise to give patients greater control over their data while improving privacy, and security. Yet these benefits are often presented as an inherent consequence of decentralization rather than as claims requiring rigorous scrutiny. This raises a fundamental question: does decentralizing identity actually improve the security and privacy of healthcare systems, or does it shift where vulnerabilities arise? To address this, we conduct a systematic literature review of recent decentralized identity architectures proposed for healthcare, and examine their security and privacy foundations. Each architecture is assessed through two threat-modeling frameworks i.e., STRIDE for security and LINDDUN for privacy, allowing us to gauge how far it addresses the threats it is intended to counter and to surface the residual risks, trade-offs, and open questions that remain.

Paper Nr:	126
Title:	A Privacy-Preserving Federated Learning Framework for Intrusion Detection in IoT Networks
Authors:	Prasanna Dhanireddy, Khushboo Kumari, Kayode S. Adewole, Andreas Jacobsson and Erdal Akin
Abstract:	he rapid growth of Internet of Things (IoT) devices in smart homes has increased vulnerability to cyber threats. Traditional centralized intrusion detection systems (IDS) raise privacy and scalability concerns due to the need for centralized data collection. This study aims to develop a Federated Learning (FL)-based IDS framework that compares the performance of Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Networks (1D-CNN), and TabTransformer (TTF) using the CICIoT2023 dataset. Differential Privacy (DP) and quantization were investigated to improve data protection and computational efficiency, respectively. TTF achieved the best detection performance with an accuracy of 93.97%, F1-score of 92.47%, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) score of 98.11%, while MLP provides superior computational efficiency. Integrating DP produces a privacy–utility trade-off, reducing TTF accuracy from 93.97% to 91.40%. Post-training dynamic quantization (PTDQ) reduces TTF and 1D-CNN model size by 67.72% and 65.66% respectively, with minimum performance loss. The optimized MLP framework effectively detects CTGANgenerated zero-day attacks with 90.96% accuracy, demonstrating high practical viability for resource-constrained IoT edge environments.

Area 4 - Human-Centric AI and Socio-Technical Information Systems

Full Papers

Paper Nr:	34
Title:	Integration of Explainable Artificial Intelligence and Human–Machine Interaction into Autonomous Shuttle Information Systems: Contextual Explanations and Passenger Experience
Authors:	Ibrahim Cherrate, Hayet Hammami and Alain Bouju
Abstract:	Autonomous shuttles are increasingly being integrated into public transportation services. They rely on an onboard mobility information system capable of collecting, processing, and exploiting real-time data in order to make automated driving decisions. However, the absence of a human driver, traditionally perceived as a source of safety in public transport, as well as the opacity of decisions produced by artificial intelligence models, may negatively affect user experience and the acceptance of autonomous shuttles. In this context, this study proposes to combine two complementary domains: human–machine interaction, which acts as a communication channel between the shuttle’s information system and passengers, and explainable artificial intelligence, which makes system decisions more transparent by generating explanations that are comprehensible to passengers. Our research focuses on the relevance of explanations related to the actions of an autonomous shuttle’s information system according to the temporal driving context, distinguishing between routine situations, such as approaching a pedestrian crossing, and critical situations, such as sudden braking caused by the unexpected appearance of an obstacle. An experiment conducted with 52 participants made it possible to evaluate the impact of these explanations on acceptance, user experience, and the way they are presented. The results show that explanations must be context-aware, must preserve the continuity of trip-related information rather than replacing it, and that the level of detail of explanations must be adapted to the situation.

Paper Nr:	113
Title:	Trustworthy AI-Augmented Information Systems through Acoustic Verification of LLM-Generated Pronunciation Feedback
Authors:	Ons Sayeh, Zeineb Fki and Boudour Ammar
Abstract:	Large language models generating explanatory text over an- alytical systems risk fluent but unsupported claims, a trust gap we study in computer-assisted pronunciation training, where an LLM explains nu- merical scores to learners. We propose a framework built around an acoustic verification gate. It checks whether each generated feedback statement matches its corresponding score, rejecting mismatches and triggering regeneration until feedback is faithful to the evidence. The framework combines an assessment model, a frozen Whisper-medium encoder with a lightweight regressor, which beats an MFCC baseline on SpeechOcean762 (PCC 0.822 vs. 0.763; p < 0.0001), with an LLM feed- back generator governed by this gate. Across three open-weight LLMs, unverified feedback contains unsupported claims in 24 to 72% of cases, unrelated to model size. Our gate raises faithfulness to 80% or higher for all models, with no retraining required. This is the first framework to treat LLM feedback faithfulness as a measured, enforced property in an educational AI system, arguing that trustworthiness, not fluency, should be a core evaluation criterion.

Area 5 - Inductive Learning, Machine Learning, and Trustworthy AI

Full Papers

Paper Nr:	56
Title:	X-EndoML: A Machine Learning-Driven Framework for Endometriosis Risk Explainable Prediction
Authors:	Anca-Elena Magui, Gabriela Czibula and Cristina Mircea
Abstract:	Endometriosis affects approximately 10% of reproductive-aged women, yet diagnosis is typically delayed by 7-10 years due to laparoscopic surgery constraints. This paper proposes X-EndoML, an analytical and predictive framework designed to validate a model-agnostic methodology for the selection, structural evaluation, and explainable prediction of relevant endometriosis symptoms. The framework employs a consensus-based filter pipeline to select 35 highly predictive symptoms from the original 58-dimensional space. The structural quality of this reduced symptom set is validated using unsupervised clustering, yielding highly coherent patient cohorts (Silhouette~= 0.555, Davies-Bouldin~= 1.122), alongside a custom quality metric that improves by +0.018 compared to the full set. An optimized, Youden-adjusted Random Forest classifier achieves an F1-score of $0.936$ on a hold-out test set, comparable to the 0.947 obtained on the full 58-symptom space. Furthermore, we leverage TreeSHAP to generate global and patient-level risk justifications, which are validated through clinical and biological interpretation.

Paper Nr:	61
Title:	When to Intervene? A Temporal Attention Framework for Early Warning Point Detection in Online Learning Systems
Authors:	Ahlem Makni, Khouloud Boukadi and Mourad Abed
Abstract:	Early detection of at-risk students remains a fundamental challenge in online learning, yet existing predictive models on the Open University Learning Analytics Dataset (OULAD) rely on arbitrarily fixed temporal windows (4, 5, or 32 weeks) without empirical justification. More critically, none address the operational question of when prediction becomes reliable enough to trigger intervention. This paper introduces a lightweight Temporal Attention Hybrid Model that integrates demographic features with weekly activity sequences (38 weeks × 20 activity types). We systematically evaluate prediction performance at selected observation weeks from week 2 to week 38 to identify the Early Warning Point (EWP)—the earliest week where prediction reliability stabilizes. Experiments on 26,074 students demonstrate three key contributions. First, our model achieves 88.2% accuracy with only 174,915 parameters, approximately 14× fewer than the 2.5M-parameter Bi-LSTM baseline. Second, we identify week 30 as the optimal EWP (87.4% F1), while acceptable reliability (85%) is reachable by week 10, offering educators a flexible trade-off. Third, attention weights reveal that weeks 16–17, 24–25, and 33 receive the highest temporal pooling importance, challenging early-window assumptions. Our framework provides intrinsic interpretability and answers the "when to intervene" question, offering a deployable architecture for cooperative human-AI decision making in educational information systems.

Paper Nr:	81
Title:	Synthetic Tabular Data Generation for Large-Scale Adult Skills Assessments: A Case Study on PIAAC Czech
Authors:	Duc Thien Pham, Roman Mouček and František Kalvas
Abstract:	In the 21st century, high-quality data is as valuable as gold for solving systemic problems; however, privacy constraints and the risk of low-quality evidence often limit its utility. Synthetic data addresses these challenges by providing a high-fidelity alternative that preserves statistical integrity while mitigating disclosure risks. This study presents a comprehensive evaluation of the PIAAC Czech Republic datasets, positioning human-centered synthetic data as a solution for large-scale adult skills assessment. We benchmark five advanced architectures: Conditional Variational Autoencoder (CVAE), Conditional Tabular Generative Adversarial Network (CTGAN), Transformer-CVAE (T-CVAE), Diffusion, and Transformer-Diffusion (T-Diff), to generate synthetic datasets across varying sample sizes (10,000 and 20,000) and feature densities (100, 300, and 500). Our results identify Diffusion as the most effective approach for high-dimensional tabular synthesis, consistently achieving a strong balance between statistical accuracy and empirical privacy metrics. These findings establish a framework for applying generative deep learning to complex social data, demonstrating that Diffusion-based synthesis preserves intricate behavioral and demographic dependencies essential for social interaction modeling, while maintaining a separation from sensitive training instances.

Paper Nr:	128
Title:	Event-Driven Real-Time Adaptation of LLM-KG-Grounded Recommendations: Application to Oncology Physical Activity
Authors:	Djebabria Hiba, Sam Yacine and Messai Nizar
Abstract:	Recommendation systems in dynamic domains face a fun- damental challenge: outputs generated from a static, first-time snapshot of user data become inaccurate. This is particularly critical in oncology, where patient clinical parameters change continuously and sometimes unpredictably during and after the acute phase. Existing healthcare rec- ommendation approaches rely on periodic recomputation and historical behavioral modeling, which don’t directly monitor the input state for clinically significant transitions. We propose an event-driven real-time adaptation framework built around three main contributions: (1) a for- mal multi-source heterogeneous state model that formalizes continuous Internet of Things vital streams, on-demand patient self-reports, and discrete clinical updates respectively, each at its own frequency, this for- malization allows us to provide a clear representation of the patient’s evolving state; (2) a per-source input-side ∆-check with a normalized deviation that selectively triggers a thorough recomputation only when a meaningful state change is detected; and (3) a Knowledge Graph thresh- old construction where safety thresholds are not hard-coded but context- aware values dynamically queried from a constructed knowledge graph as a function of the current patient state. Experiments conducted on 30 patient profiles demonstrate that our approach removes violations of contraindications present in the static baseline, yet manages to detect clinically significant state transitions overlooked by the output-side un- certainty trigger compared with two chosen baselines.

Area 6 - Intelligent Process Analytics and AI-Driven Automation

Full Papers

Paper Nr:	38
Title:	Do My Log and Model Originate from the Same Process? Log-Model Hypothesis Testing
Authors:	Felix Britzelmaier, Cameron Pitsch and Sander J. J. Leemans
Abstract:	Process mining helps organisations understand and improve their processes by analysing recorded event data. A core capability is comparing event logs against process models to identify how observed behaviour diverges. Yet when log and model differ, current techniques quantify the difference but cannot tell whether it reflects a ``real'' deviation or the variability expected in a finite log of a stochastic process. Process hypothesis testing fills this role by computing a p-value, but existing tests either only compare two event logs or address the log-to-model setting with increased Type 1 error rates. We propose a test for the hypothesis that an event log and a stochastic model originate from the same distribution. The test iteratively draws independent samples from the model, runs a permutation test against the log, and aggregates the results into a single p-value, accommodating models with infinite trace languages. This extends the methodology for permutation based hypothesis testing in process mining to the log-to-model setting. Evaluation shows that the Type 1 error rate is well controlled at conventional significance levels, achieves detection accuracy competitive with existing techniques, and reflects model quality consistently with stochastic conformance metrics.

Paper Nr:	46
Title:	DVIDEO: Detecting and Visualising Quality Issues in Object-Centric Event Logs
Authors:	Florian Eckhard, Jan Krückemeier, Holger Wittges and Stefanie Rinderle-Ma
Abstract:	Process mining is a data-driven method for identifying processes as they occur. Object-centric process mining has emerged as an approach that provides a more holistic and realistic view of complex processes. However, this approach results in more complex event logs that require greater preprocessing and modelling effort. To support analysts during preprocessing, we propose 11 object-related issue detection patterns and four visualisation patterns, specifically targeting objects and object-to-object relations as the core differentiators from case-centric process mining. We define the patterns, implement them, and evaluate their performance. The detection pattern evaluation shows a median detection rate of 85%. For the visualisation, we conducted a quantitative survey (n = 24) measuring task success rate (83%), perceived usefulness (93% positive impact) and perceived ease of use (>80% for all items). For future research, we recommend applying our patterns to further quality issues and evaluating our approach in a practical project.

Paper Nr:	49
Title:	Balancing Trustworthiness and Flexibility for LLM-Based Process Compliance Pattern Extraction from Text
Authors:	Johannes Loebbecke, Juergen Mangler and Stefanie Rinderle-Ma
Abstract:	Business process compliance (BPC) aims to ensure that processes adhere to legal, regulatory, and internal constraints. Violations may cause financial or reputational damage, prompting BPM researchers to develop verification approaches applied before, during, or after process execution. Regardless of verification time, BPC approaches require extraction of machine-processable constraints from natural language documents such as regulatory documents. Large Language Models (LLMs) show promise for this task and have been used to directly verify process models or to formalize compliance constraints as logical formulae. However, direct verification is prone to hallucinations and lacks trustworthiness, while logic formulae require expert validation and custom elements. Hence, we propose an LLM-based approach for extracting composable, compliance patterns from text that encode compliance properties as Abstract Syntax Trees (ASTs), contributing to the flexibility of the approach. The resulting human- and machine-readable ASTs are compatible with various BPC approaches via formal mappings, ensuring trustworthiness through model checking. The approach is prototypically implemented and evaluated through a user study comparing human and generated encodings, an empirical analysis of input/generation parameters, and a higher-education case study assessing real-world applicability.

Paper Nr:	79
Title:	Adaptive BPM Systems: Integrating Process Mining and Multi-Objective Reinforcement Learning for Preference-Aware Decision Support
Authors:	Bupe Chindongo and Hendrik Johannes Carl van der Westhuizen
Abstract:	Modern business process management (BPM) systems often treat process optimisation as a static, design-time activity, leaving organisations unable to respond to shifting cost, time, and compliance priorities without costly reconfiguration. Existing multi-objective methods generate process trade-offs offline and must be re-run when priorities move, whilst reinforcement learning applied to business processes typically optimises a single scalar objective. We contribute an adaptive BPM information system, built on multi-objective reinforcement learning, in which the objective weighting rather than the process model is what changes at runtime. It discovers a Petri net from event logs, derives cost, time, and compliance objectives from that same log, and computes vector-valued Q-functions over them by dynamic programming, so that any stakeholder preference weight is served without retraining. Case-matched evaluation against historical baselines on 1,000 real-world financial approval cases demonstrates Pareto improvements under balanced preferences (normalised-scale deltas of 0.305 cost, 0.300 time, 0.947 compliance, all p < 0.0001). Cost-time prioritisation yields larger efficiency gains (0.435 cost, 0.441 time, p < 0.0001), whilst a nonsignificant aggregate compliance delta (−3.2%) masks opposing effects by case value: gains for lower-value cases and losses for higher-value ones, tracking amount-dependent approval requirements. A trade-off that remains implicit in a fixed process design, therefore, becomes explicit and quantifiable per case segment, letting process owners weigh operational gain against regulatory risk as priorities shift.

Paper Nr:	85
Title:	Miss Probabilities for Subprocesses
Authors:	Philipp Bär, Adam T. Burke, Moe T. Wynn and Sander J. J. Leemans
Abstract:	Conformance checking compares observed and modeled process behavior to explain where reality deviates from a process model. For targeted process improvement, applications such as model repair and process comparison benefit from stochastic guidance alongside such explanations: Likely deviations are typically prioritized over edge cases. Existing measures quantify how likely parts of a modeled process are superfluous in reality, but lack the converse dimension, namely how likely observed reality is missing from the model. Consequently, phenomena such as rework may remain hidden operational bottlenecks. In this paper, we introduce miss probabilities for process trees, which stochastically quantify behavior that is observed in reality but that is absent from the model. Based on all optimal explanations for deviations, we infer which parts of a process model lack which observed behavior and formally quantify this lack as a miss probability. That way, miss probabilities yield localized diagnostics for conformance analysis and process analytics. We empirically evaluate miss probabilities on real-life processes and outline their potential for process improvement. Our experiments show that miss probabilities reveal local incompleteness of the model even in highly concurrent settings and provide insights for model repair.

Paper Nr:	103
Title:	Data-Driven Heuristic-Guided Business Process Redesign
Authors:	Nan Sai, Karolin Winter and Remco Dijkman
Abstract:	Business Process Redesign (BPR) has a long-standing catalog of best-practice redesign heuristics, yet, applying them remains a mostly manual task that requires extensive domain knowledge. Despite recent advances using Large Language Models (LLMs) for conversational process model redesign, limitations remain. For example, redesign is user-driven instead of data-driven and is not guided by best practices such as the redesign heuristics. Moreover, only one change can be handled at a time. To address these limitations, given an event log and a BPMN model, we present the DH-BPR framework to realize data-driven heuristic-guided BPR using LLMs. To design DH-BPR, we first conducted an analysis of which redesign heuristics can be automatically detected and applied given information contained in an event log such as the fraction of cases that bypass an activity. Based on this, DH-BPR detects and validates heuristic candidates through a combination of rule-based techniques and an LLM. After that, DH-BPR jointly designs and records an application plan for valid candidates, before finally yielding a redesigned BPMN model and an updated resource model (the way resources execute the process). Simulation-based evaluation on a synthetic benchmark and two real-world logs shows that the framework's redesign reduces mean throughput time per case by up to 61.5%.

Short Papers

Paper Nr:	26
Title:	FINTR: Optimization-Based Integration of HIS and RTLS Data for Nursing Task Retrieval
Authors:	Haroon Tharwat, Niels Martin and Benoît Depaire
Abstract:	Process mining offers great potential to support the improvement of healthcare processes, but the Hospital Information Systems (HIS) data that is often used suffers from data quality issues such as delayed or missing task entries. Real-Time Location Systems (RTLS) provide an alternative data source as they capture staff and equipment positions over time, providing more accurate timestamps, but do not contain semantic information on the tasks performed. While there are clear synergies between both data sources and both have been considered in isolation in process mining research, a method to systematically integrate them to obtain an enhanced input for process mining is missing. This paper introduces FINTR (Footprint-based Integration for Nursing Task Retrieval), a semi-automated method that integrates HIS and RTLS data via task footprints (templates that encode how each nursing task type should appear across both data sources) to provide an enhanced foundation for process mining in healthcare. Preliminary evaluation on synthetic data shows that FINTR reconstructs nursing tasks in a noise-free setting and reveals how its performance evolves when increasing levels of a common type of noise in real-world settings are present.

Paper Nr:	31
Title:	Exploring Complexity-Related Failure Patterns in LLM-Generated BPMN 2.0 Process Models
Authors:	Luca Franziska Hörner, Maximilian Möller and Manfred Reichert
Abstract:	Recent advances in Large Language Models (LLMs) enable automated BPMN 2.0 process model generation from natural language descriptions. However, the relationship between process description complexity and generation quality remains insufficiently understood. To address this gap, we conducted a controlled empirical evaluation using BPMNGen, our framework for generating, validating, and analyzing LLMgenerated BPMN 2.0 process models. Based on 100 generated process models, we investigated the relationships between linguistic and structural complexity, recurring failure patterns, syntactic correctness, semantic alignment, and generation efficiency. Linguistic complexity was mainly associated with sequence flow-related failures, reduced semantic alignment, and longer generation times, whereas structural complexity showed broader associations with sequence flow inconsistencies and deadlocks, lower syntactic correctness, reduced semantic alignment, longer generation times, and higher output costs. Across the evaluated LLMs, deadlocks and sequence flow-related failures represented the main recurring failure patterns, although overall failure frequencies remained low.

Paper Nr:	40
Title:	A Matter of Perspective: Case Notion Selection for Object-Centric Performance Analysis
Authors:	Jan Niklas van Detten, Pol Schumacher and Sander J.J. Leemans
Abstract:	Information systems track the interactions between business objects, which can be extracted into object-centric event logs. Object-centric process mining aims to analyze these logs to ideally improve relevant performance measures. This analysis requires a clear notion of which objects belong to a joint process execution, commonly called a case. To produce relevant analytical results, such a case notion needs to be carefully chosen. This poses a challenge in practice due to the exponential number of case notions to choose from. In this paper, we introduce a novel framework to semi-automatically select case notions for the performance analysis on object-centric logs. Manual work by practitioners is reduced to selecting a performance measure and a distribution property that indicates optimization potential. We evaluate our approach on ten public logs and find it to produce relevant insights in feasible runtime.

Paper Nr:	51
Title:	Semantic Coherence Drifts in Process Models: Iterative Reverse Transformation for LLM-Driven Conversational Process Modeling
Authors:	Nataliia Klievtsova, Juergen Mangler, Daniel Schuster, Timotheus Kampik and Stefanie Rinderle-Ma
Abstract:	Large Language Models (LLMs) enable the automatic creation of process models from text. These models can be iteratively refined in conversations, where the LLM repeatedly takes its own prior output as input. Preserving semantic coherence throughout the conversation is crucial, as drift compounds with each refinement request and can lead to unexpected results. Existing quality metrics (e.g., accuracy, hallucination, bias, fairness, reliability) typically require ground truth, i.e., knowledge of the expected output of a refinement request, which is often unavailable or infeasible to obtain. We therefore investigate iterative reverse transformations (model-text-model loops) to automatically detect such semantic coherence drift in LLM-driven process model creation. This lets us pinpoint when drift occurs and how it evolves over iterations, enabling comparison of LLM and pipeline reliability and realistic expectations for their iterative use. The approach is tested on several data sets, including both artificial and real-world process models.

Paper Nr:	89
Title:	Towards Explainable and Interpretable Predictive Monitoring of Collaborative Processes
Authors:	Roberto Nai, Emilio Sulis, Daniel Calegari, Andrea Delgado, Guido Boella and Adriana Boccuzzi
Abstract:	Collaborative processes involve multiple interacting participants whose coordination and interaction dynamics make predictive process monitoring (PPM) particularly challenging. Moreover, many predictive approaches remain black-box models, limiting their interpretability. This work proposes integrating explainability and interpretation components into a PPM pipeline for collaborative business processes. The approach enriches traditional event logs with participant- and interaction-aware information. The pipeline combines Transformer-based prediction, SHAP explanations of activity- and participant-related contributions, and a locally deployed Large Language Model for privacy-preserving human-readable interpretation. The evaluation considers four benchmark collaborative event logs and a new real-world Emergency Department event log.

Paper Nr:	105
Title:	Planning and Scheduling Business Processes under Control-Flow Uncertainty
Authors:	Michel Kunkler and Stefanie Rinderle-Ma
Abstract:	Scheduling activities in business processes can improve effi- ciency, but is challenging because the sequence of activities needed to complete a case is often uncertain, as it depends on data revealed dur- ing execution. Probabilistic information on these decisions can, however, be derived from historical execution logs. Planning with particular paths trades off feasibility, i.e., the probability of successful completion, against the expected number of superfluous activities that are planned but never executed. We frame this as a chance-constrained optimization problem and present two formulations: a decomposed approach that minimizes superfluous activities subject to a feasibility constraint before minimiz- ing the makespan, and an integrated approach combining both stages. Evaluation on two real-world and one synthetic dataset shows that the integrated approach yields superior makespans but is intractable at scale, while the decomposed approach scales to large settings.

Paper Nr:	109
Title:	Assessing Resource Availability for Business Process Simulation Calibration
Authors:	Mei Wang, Karolin Winter, Quang-Vinh Dang and Remco Dijkman
Abstract:	Business process simulation models are commonly assessed by comparing simulated and observed event logs through aggregate performance measures. Such comparisons show whether the simulation behavior matches real-world behavior, but not which simulation model settings should be revised to improve that correctness. This also applies to resource availability settings, which determine when resources work. Assessing these settings from event logs is difficult, because event logs record when work actually started, but not when it could have started. A working period without a recorded start may mean that the resource was actually unavailable or could have started work but simply did not. To address these problems, this paper presents an approach for assessing resource availability settings from observed and simulated event logs. The approach calibrates the resource availability setting by changing selected availability probabilities. The evaluation examined whether calibration made the behavior produced by the simulation model more consistent with the observed log. It showed that this was the case in both evaluations on synthetic and real-world event logs.

Paper Nr:	110
Title:	Decision-Aware Suffix Prediction and Reasoning of Business Processes
Authors:	Henryk Mustroph and Stefanie Rinderle-Ma
Abstract:	Suffix prediction forecasts the remaining sequence of events of a running case until completion. Most approaches rely on neural networks (NNs) trained on event logs, which, on average, perform well but struggle with short prefixes or targets belonging to a rare process variant. In such scenarios, the correct path may cross multiple branching decisions, determined primarily by case- and event-level attributes, a signal that NN-based suffix prediction models tend to underweight because they may heavily weight (dense) event labels. Decision mining extracts rules for such decisions from the event log, but has so far been applied only to post-hoc and what-if analysis, not suffix prediction. We therefore extend suffix prediction with decision mining, introducing a decision-aware suffix prediction framework, a neuro-symbolic approach that enables reasoning about predicted events via mined decision rules. Experiments on three of four event logs and three suffix predictors show that the framework can improve suffix prediction, especially for short prefixes but also for rare process variants, and adds intrinsic interpretability.

Paper Nr:	112
Title:	Intersections as Cooperative Information Systems: V2X-Driven Reinforcement Learning for Emergency-Aware Traffic Control
Authors:	Hamdi Mnif, Hend Marouane, Mohamed Mosbah and Hassene Mnif
Abstract:	Urban traffic congestion and emergency response delays remain major challenges for intelligent transportation systems, directly affecting mobility, public safety, and service efficiency. Traditional fixed-time traffic signal controllers are limited when facing non-recurrent events such as emergency vehicle arrivals, since they cannot proactively adapt signal phases while preserving regular traffic flow. The proposed approach models the intersection as a cooperative information system in which vehicles, emergency vehicles, roadside units, and the traffic signal controller exchange information through V2X-inspired messages, including signal phase information and priority requests. Within a cooperative architecture, a Proximal Policy Optimization agent is used as the adaptive decision engine to select signal phases according to queue conditions, current signal phase, elapsed phase time, phase-change readiness, and emergency vehicle presence. An emergency-first reward function is designed to prioritize emergency vehicle clearance while reducing average waiting time and queue imbalance for regular vehicles. The framework is implemented in a custom SUMO–Gym environment and evaluated on a four-arm signalized intersection under different traffic demand levels. Compared with a fixed-time baseline, the proposed controller reduces average vehicle waiting time by 58.6% in normal traffic and emergency vehicle waiting time by 69% in the emergency scenario. A runtime feasibility analysis further shows that the PPO decision latency remains lightweight, with an average measured latency of 24.13ms.

Area 7 - Internet of Things, Cyber-Physical Systems, and Digital Twins

Full Papers

Paper Nr:	68
Title:	SafeOrchestrate: Runtime-Checked Intent-Driven Orchestration for Cyber-Physical Systems
Authors:	Muhammad Ashfaq, Ahmed R. Sadik, Ewan Warburton, Pyry Kotilainen, Niko Mäkitalo and Tommi Mikkonen
Abstract:	Modern cyber-physical systems (CPS) increasingly integrate heterogeneous sensors, actuators, controllers, and services into compositions formed at runtime. Natural-language (NL) orchestration of such systems is attractive, but unconstrained large language model (LLM) tool use is risky because actions must respect device capabilities, runtime state, and safety constraints. We present SafeOrchestrate, a reference architecture for runtime-checked, intent-driven CPS orchestration. Given an NL intent, SafeOrchestrate decomposes it into atomic goals, generates a runtime-checkable specification for each goal, grounds these specifications in discovered device affordances, and gates candidate actions against the runtime state. For single-action consequents, it also checks each action against the complete grounded specification and any safety constraints before execution. In our Web of Things (WoT) instantiation, SafeOrchestrate produces Linear Temporal Logic (LTL)-style specifications, grounds them in Thing Description (TD) affordances exposed through the Model Context Protocol (MCP), and enforces satisfiability modulo theories (SMT)-backed guardrails on single-action consequents. We evaluate this instantiation against a direct LLM-to-MCP baseline on simulated smart home, greenhouse, and industrial automation tasks. Cross-rule and safety-constraint checks are illustrated separately through a worked example. SafeOrchestrate achieves 100% required-action accuracy on every task and 100% MCP Execution Similarity (MCP-ES) on all but one, where MCP-ES is 92%. By contrast, the baseline degrades on multi-device, event-triggered, and ordered tasks, with MCP-ES falling to 13%. On these tasks, the results suggest that combining LLM-based specification generation with role-aware grounding and runtime constraint checking improves reliability for open CPS orchestration without enriched TD annotations or planning-domain expertise.

Short Papers

Paper Nr:	121
Title:	A Digital Twin-Enabled Middleware for Smart Contract Lifecycle Management in Blockchain-Based Business Processes
Authors:	Ryhem Essid and Hatem Hadj Kacem
Abstract:	Smart contracts offer a credible path to automating inter- organizational business processes, but deployment reveals coordination gaps that existing tools do not bridge. Generating contract code from a BPMN model handles only the static part of the problem; keeping that contract synchronized with a live process engine, streaming IoT data, and predictive outputs at runtime remains open. We present a middleware that closes this gap through a Digital Twin, a continuously updated replica of the running processes and their physical environ- ment, driving three tiers responsible for communication, runtime control, and governance. Rather than acting as a passive visualization layer, the Digital Twin ingests sensor streams, queries a machine-learning service, and propagates verified decisions to both the process engine and the blockchain. Feasibility is shown through an agricultural irrigation pro- totype built on Camunda 8, Zeebe, and an Ethereum-based blockchain, with a request-to-instantiation latency of 79 ms and per-run contract de- ployment of 93 ms on a local node, and full on-chain traceability under nominal and degraded conditions. The design generalizes beyond agri- culture to any BPMN–IoT–blockchain process.

Area 8 - Knowledge Graphs, Data Engineering, and Intelligent Knowledge Systems

Full Papers

Paper Nr:	58
Title:	Conflicted Records: Typed Adversarial Contracts for Auditable LLM Output Quality Enforcement
Authors:	Kardelen Subasi and Jarosław A. Chudziak
Abstract:	LLM-generated records increasingly populate production information systems, treated as data assets subject to the same quality expectations as records from operational databases. A single such record can simultaneously satisfy and violate a quality rule: a grounded fact next to a hallucination, a useful answer next to a stereotype. Classical data cleaning cannot resolve this, because the two judgments are incompatible over the same span, and a single LLM judge cannot either, since it shares the training biases that produced the conflict. Prompt engineering, chain-of-thought, and RLHF lower surface error rates but leave no audit trail and cannot distinguish genuine remediation from blanket suppression. The question is whether a deterministic check over structured inter-agent records can catch the defects that referential checks alone miss. \\ We present BiasGuard-D: two adversarial roles drawn from distinct model families exchange typed claim documents, a three-model quorum adjudicates by majority vote, and a combined referential-plus-semantic invariant validates the verdict. The semantic component fires on verdicts referential checks miss; removing the iterative loop drops invariant satisfaction sharply. The deliberation layer transfers from stereotype to hallucination remediation on HaluEval QA with only role-framing changes, raising right-answer recovery by 49 percentage points, evidence that adversarial cross-family deliberation is a reusable defect-remediation primitive rather than a stereotype-specific fix; a cross-family second auditor is required to maintain recall on objective hallucination. Every claim, vote, and instruction is stored as a typed record, giving operators a provenance trail auditable under emerging AI regulations.

Short Papers

Paper Nr:	57
Title:	Wikis That Disagree: Pluralism as a Representation Primitive for Deliberative Knowledge Systems
Authors:	Yousef Attia and Jarosław A. Chudziak
Abstract:	Cooperative information systems increasingly rely on large language model (LLM)-curated shared knowledge bases. In pluralistic domains, stakeholders share the same evidence but reach different conclusions; consolidating their positions into one record destroys the attribution distinguishing them, and no downstream layer can reconstruct what curation discarded. We ask whether encoding pluralism as a schema constraint on every record, rather than as routing logic over anticipated questions, changes what a system can faithfully report. We present Mushawara: an LLM-curated wiki for Islamic jurisprudence whose schema mandates a parallel track per school of law on every page, built by a deterministic ingest pipeline and queried by a five-agent system. It attributes and cites substantially more answers than every baseline tested, including a graph-augmented retrieval variant, with accuracy comparable across conditions. Attribution quality is bounded by representation quality, not reasoning quality, a pattern any domain with shared evidence and divergent stakeholder conclusions can reuse.

Paper Nr:	123
Title:	A Semantic-Cooperative Decision-Support Architecture for Explainable Demand-Responsive Transport Management
Authors:	Nourelhouda El Gadi, Esma Talhi and Alain Bouju
Abstract:	Demand-responsive transport (DRT) systems are increasingly important in urban mobility, especially under dynamic demand, operational constraints, and network disruptions, which require coordination across users, vehicles, infrastructure, and external events. Existing approaches to the Dynamic Dial-a-Ride Problem (DDARP) provide effective optimization solutions but rely mainly on numerical network representations, thus offering limited support for interpreting the operational context of their decisions. This paper proposes a knowledge-based cooperative information system combining a semantic knowledge base with an ontology-driven multi-objective optimization module. The ontology integrates mobility-related information, including demand, GTFS-based public transport, infrastructure, spatio-temporal knowledge, and disruptions. Beyond its descriptive role, it actively supports the decision-making process by using semantic reasoning to enrich the optimization model with domain knowledge and provide explicit explanations for the resulting decisions.

Area 9 - Services, Cloud in Information Systems and Distributed AI Platforms

Full Papers

Paper Nr:	70
Title:	Q-PARC: A Pressure-Aware Resource Controller for Hadoop YARN Queues
Authors:	Helmi Tlich, Hanène Chettaoui and Tarek Hamrouni
Abstract:	Shared information-system platforms increasingly rely on distributed data-processing backends where multiple services, tenants, and analytical workloads compete for resources. In Hadoop YARN, this sharing is often managed through CapacityScheduler queues whose capacities and limits remain static, even when runtime congestion and I/O pressure change. This paper presents Q-PARC, a pressure-aware queue controller for shared YARN-based information-system platforms. Q-PARC runs as an external control layer, observes queue and node metrics at fixed windows, separates queue congestion from I/O pressure, and applies bounded updates to queue shares, per-queue memory limits, and concurrency caps. The controller is evaluated on an 11-node Hadoop cluster using WordCount and TeraSort workloads under congestion burst, I/O burst, and mixed scenarios. Compared with Static YARN, CapacityScheduler with preemption, and a congestion-only controller, Q-PARC provides the best congestion/I/O trade-off. It reduces waiting time in congestion and mixed scenarios, remains close to the best method in the I/O burst scenario, and achieves the lowest spill volume and local I/O failure rate. These results show that lightweight queue-level resource orchestration can improve the operational behavior of shared distributed information-system platforms without replacing the native scheduler.

Paper Nr:	76
Title:	Reducing Service API Schemas for LLM-Based Text-to-GraphQL
Authors:	Yann Ramusat, Charles Prioux, Jeremy Mechouche and Zakaria Maamar
Abstract:	As GraphQL becomes a pivotal technology for service composition and orchestration in distributed architectures, Large Language Models show strong potential for automating service discovery via Textto-GraphQL generation. However, enterprise-grade service ecosystems expose massive, strongly-typed API contracts (schemas) that easily exceed the token context limits of these models. Existing approaches, such as Retrieval-Augmented Generation and iterative agentic exploration, often fragment the underlying service graph structure, leading to structural hallucinations. We propose a general-purpose framework for topological schema reduction, formulating service schema navigation as constraintpreserving subgraph extraction. The framework employs a dependencyaware breadth-first search to construct a Minimal Schema Definition Language (SDL)—a compact, self-contained subgraph encapsulating the exact business entities and structural wrappers required to fulfill user intent. Evaluated across 14 APIs, including production-scale schemas up to 4k types such as GitHub and GitLab, the framework achieves up to 96.1% input token reduction while preserving topological integrity. Empirical results demonstrate a 66.8% extraction success rate, significantly outperforming the 11.8% achieved by rule-based iterative baselines. In end-to-end generation, the Minimal SDL yields a 53.2% query validation success rate (compared to a 40.7% full-schema baseline), effectively mitigating context overflow and proving its viability for AI-driven service orchestration.

Paper Nr:	77
Title:	Improving MCP-Based LLM-Service Integration with Enriched API Documentation
Authors:	Ada Bagozi, Devis Bianchini, Massimiliano Garda, Michele Melchiori and Stefano Picchi
Abstract:	In recent years, Large Language Models (LLMs) have increasingly been used as conversational intermediaries between users and service APIs, mapping natural language requests to API invocations. In this paradigm, API endpoints are exposed to LLMs as callable tools derived from API documentation. However, when such documentation fails to capture execution behavior, LLMs may generate incorrect or incomplete tool invocations, leading to execution errors. This issue is further exacerbated when handling complex user requests, which often require coordinating multiple tool invocations. In such cases, LLMs must not only identify the appropriate tools but also generate a correctly ordered sequence of invocations (i.e., a tool plan) while maintaining consistent parameter propagation across successive calls. To address these challenges, this paper proposes an approach to LLM-Service integration for dependable service orchestration and execution based on the Model Context Protocol (MCP). Specifically, the approach combines: (i) a service modeling framework that enriches standard OpenAPI documentation with properties (including intents, preconditions, effects, produces/consumes descriptors, and dependency relationships) directly interpretable by LLMs to support tool selection and stepwise plan construction; (ii) a tool plan construction and execution process, in which the MCP client uses the enriched documentation to support on-the-fly tool plan generation, context-aware parameter propagation, and humanin-the-loop supervision for sensitive state-changing operations. Experimental results on a real-world retail service ecosystem, obtained through an ablation study, provide initial evidence of the impact of enriched service documentation on tool plan construction and execution quality.

Short Papers

Paper Nr:	27
Title:	MAPS: Microservice Adaptive Placement Strategy in Kubernetes
Authors:	Andreas Lopata and Euripides G. M. Petrakis
Abstract:	Kubernetes (k8s) automates the deployment, scaling, and management of containerized applications. A K8s cluster consists of Nodes (VMs) that can host one or more microservices (Pods). Scheduling (i.e., placing Pods into VMs) becomes a complex problem for applications with high-affinity microservices that exchange high traffic. High-affinity microservices must be grouped and placed in the smallest possible number of VMs to maximize resource utilization and minimize application latency and Egress traffic. In this work, the problem of placement is modeled as graph clustering. MAPS traces workload changes and applies an adaptive density-based clustering method to continuously update microservice placement on Nodes, making it suitable for dynamic microservice placement under fluctuating workloads.

Paper Nr:	130
Title:	Temporal JSON Query
Authors:	Daniel Lemos de Morais and Curtis Dyreson
Abstract:	JSON Query is a language for querying data formatted in JavaScript Object Notation (JSON). But JSON Query, like all other JSON query languages, can only query the current snapshot of data. JSON documents evolve over time as data is inserted, updated, or re- moved, and many applications need to reason about the evolution of the data. Temporal JSON is a representational model of the evolving history of a JSON data collection, in which time metadata is interleaved with JSON data. This paper introduces Temporal JSON Query (TJQ), which adds temporal support to JSON Query (JQ). In TJQ a user can navigate past states and versions of the data and evaluate sequenced JQ queries; a sequenced query is one in which the query is, in effect, evaluated for all states in a time frame.

Area 10 - Software Engineering for AI Engineering for Information Systems

Full Papers

Paper Nr:	48
Title:	A Gap Analysis of Multi Agent Systems and Augmented Reality for ECU Integration Testing in Automotive
Authors:	Alexander Bartik and Pascal Hirmer
Abstract:	Modern SDVs contain an ever-increasing number of software components that run on multiple high-performance computers. Even though this leads to a reduction in the actual number of ECUs, the complexity of the software continues to increase as a result of this reduction. Existing testing approaches face challenges regarding scalability, fault diagnosis, traceability, and the interpretation of large volumes of test data. While multi agent systems support distributed fault detection and automated analysis, augmented reality technologies improve visualization and user interaction in complex technical environments. However, both technologies are currently investigated largely independently in automotive research. Our paper presents a systematic gap analysis of multi agent system and augmented reality based approaches for automotive electronic control unit integration testing using the PRISMA methodology. The analysis identifies significant research gaps regarding the integration of automated analysis, transparent diagnostics, visualization, and human centered interaction within testing workflows. Our paper outlines a first approach for combining multi agent systems and augmented reality to improve transparency, usability, traceability, and diagnostic efficiency in future automotive testing environments.