CoopIS-ERA 2025 Abstracts

Area 1 - Early Research Achievement
Short Papers

Area 1 - Early Research Achievement

Short Papers
Paper Nr: 5
Title:

Towards Declarative Process Execution: Functionalities and Architecture
Authors:

Johannes Loebbecke, Dominik Voigt, Juergen Mangler and Stefanie Rinderle-Ma
Abstract: Flexible, trusted, and compliant service orchestration, where executions adhere to constraints imposed by legal or business requirements, has become paramount for organizations. Using Business Process Management Systems (BPMS) to orchestrate services that implement human and computerized tasks can increase trust, enable efficient communication, and lead to flexible and transparent service orchestration. Existing BPMS generally use the imperative paradigm, which is suited for orchestrating strict processes with few deviations, while declarative models have been proposed for orchestrating processes with many and deeply diverging paths. However, no existing process execution engine supports the declarative paradigm while preserving the advantages of BPMS. We propose a list of functional requirements of a declarative engine that can be achieved through loosely coupled integration with existing BPMS as a path towards declarative process execution.
Paper Nr: 6
Title:

Attention-Based Connectivity Matrix Computation for Autism Spectrum Disorder Diagnosis from fMRI Data
Authors:

Sirine Sboui, Mohamed Djallel Dilmi and Faten Chaieb-Chakchouk
Abstract: This paper introduces a novel approach for autism spectrum disorder (ASD) classification from resting-state fMRI by replacing conventional Pearson correlation connectivity matrices with attention-based representations. From a theoretical standpoint, we demonstrate that Pearson correlation is a special case of the scaled dot-product attention mechanism under fixed projections and normalization, which allows attention to be viewed as a strict generalization of functional connectivity. Building on this formulation, we propose an end-to-end pipeline in which attention layers directly learn connectivity patterns from fMRI time series, eliminating the need for predefined functional correlation matrices. Experiments on the ABIDE-II dataset show that attention-based connectivity significantly improves diagnostic performance across diverse architectures, including MLP, vanilla Transformers, BrainNetCNN, and the Brain Network Transformer (BNT). The proposed method achieves a state-of-the-art accuracy of 90.4% with BNT, substantially outperforming correlation-based baselines. These results highlight the potential of attention mechanisms to enhance both the accuracy and interpretability of functional connectivity modeling in neuroimaging.
Paper Nr: 9
Title:

Knowledge-Based, Context-Aware AI for Accurate Learner Classification with Resampling Techniques in Low-Resource Environments
Authors:

Sameher Ajili, Rym Cheour, Mariem Abid, Mouna Baklouti and Richard Hotte
Abstract: AI-driven prediction models face significant challenges in low-resource educational contexts, where cold start conditions and severe class imbalance data are prevalent. This study addresses this problem by proposing a methodological framework that integrates NLP-based preprocessing for feature extraction, advanced imbalance- aware resampling, and a comprehensive multi-model comparison. We evaluated the framework using a small, authentic dataset of learners from Mali, applying ten classifiers, including SVM, Random Forest, XGBoost, and ensemble methods. Model performance was assessed using stratified 10-fold cross-validation, with and without resampling via Random Oversampling (ROS) and SMOTE strategies, with evaluation based on accuracy, F1-score, and AUC-ROC. Results show that ensemble methods, specifically Random Forest, Gradient Boosting and XGBoost, achieved over 90% accuracy, and superior F1-scores, after SMOTE, significantly outperforming baselines and ROS.
Paper Nr: 10
Title:

Lightweight LoRa Signal Quality Classification Using RSSI and SNR
Authors:

Boubaker Abdallah, Rym Cheour, Jalel Ktari and Olfa Kanoun
Abstract: LoRa/LoRaWAN enables long-range and low-power connectivity but face significant challenges from signal fading, interference, and heterogeneous deployment. This paper presents supervised learning framework for classifying LoRa signal quality using non-PHY measurements (RSSI/SNR) drawn from IEEE Dataport. We introduce a four-level quality taxonomy and a reproducible pipeline with feature leakage-safe preprocessing and class-weighted imbalance mitigation. We benchmark representative classifiers, namely Logistic Regression, support vector machines with radial basis kernel, Random Forest, and a compact multilayer perceptron, under stratified five-fold cross-validation. Evaluation relies on macro-F1 as the primary endpoint, complemented by per-class precision, recall, and F1 as well as normalized confusion matrices. Error analyses highlight that residual confusions occur mainly between adjacent regimes, consistent with overlapping RSSI and SNR bands. The results indicate that lightweight models trained on RSSI and SNR can deliver robust, deployment-ready quality classification without access to physical-layer waveforms, supporting quality-of-service monitoring and adaptive configuration at LoRa gateways.
Paper Nr: 11
Title:

A Hybrid Dynamic Knowledge Graph Building Approach
Authors:

Sana Ben Abdallah Ben Lamine, Redouane Bouhamoum and Hajer Zghal
Abstract: Interest is increasingly focused on Knowledge Graphs (KGs) thanks to their ability to structure heterogeneous information into an organized and interconnected form, enabling better reasoning, decision-making, and supporting advanced AI applications. The construction method of these KGs is just as important, as it has a direct impact on automating their creation, guaranteeing their quality and their adaptation to constantly evolving needs and applications. In this paper, we propose a hybrid knowledge graph construction approach that combines the structured, ontology-driven top-down approach with the data-driven, iterative bottom-up approach, combining the rigor of top-down ontology design with the adaptability of bottom-up data integration. Our approach is three-phase : the first is a top-down ontology construction phase, the second is a bottom-up data integration and enrichment phase and the third is a hybrid integration and continuous evolution phase. Our aim is to create a more flexible construction method that takes into account structured, semi-structured and unstructured data, offering a comprehensive and dynamic KG. This approach is particularly effective for domains requiring both structured knowledge and dynamic data integration, from social media for example, like the healthcare domain. A detailed case study in this domain illustrates our proposal and demonstrates its feasibility and effectiveness.