Higher Institute of Food Industries of Tunis

Our paper evaluates lightweight Federated Learning (FL) models for privacy-preserving Arabic sentiment analysis. We propose a CPU-efficient FL framework using classical models (SVM, Random Forest, Logistic Regression, etc.) as a practical alternative to resource-intensive transformers. tested on Arabic tweet datasets (ArSAS, ASTC, SS2030), Federated SVM achieved top accuracy (e.g., ~90% on ArSAS), while Logistic Regression offered an optimal balance of high performance and minimal execution time. Results confirm FL provides comparable accuracy to centralized training with no significant performance trade-off, while ensuring data privacy, low communication costs, and scalability, making it ideal for edge devices.

This study explores tomato agri-food residues as sources of bacteria with bioprotective potential to enhance product shelf-life and safety. A total of 245 bacterial strains were isolated, comprising predominantly Pseudomonas (52%) and Bacillus (44%) spp., with lactic acid bacteria (LAB) present at lower levels (4%). The antimicrobial activity of these isolates was assessed against pathogenic and spoilage bacteria and phytopathogenic molds. Notably, the Bacillus isolate TRB1-7 exhibited moderate activity against L. monocytogenes (inhibition halo diameter: 10.64 mm), while Pseudomonas and LAB isolates showed limited or no inhibition. Antifungal assays highlighted significant antifungal potential for Bacillus isolates. Results showed that 16% and 15% of the 245 isolates inhibited F. oxysporum and C. acutatum growth, respectively. Nine of these isolates underwent acid-adaptation and were evaluated against the selected molds using Potato Dextrose Agar (PDA) at pH 4.0 to simulate tomato conditions. Only isolate BRZ3-2, identified as B. aerius, was adapted to acidic conditions and inhibited F. oxysporum by 25%. Experiments on tomato-based agar at the same pH showed no inhibition by Bacillus isolates. These results suggest that tomato microbiota harbors acid-tolerant Bacillus strains with potential for post-harvest bio-preservation. Further studies on strains TRB1-7 and BRZ3-2 are required to develop effective bioprotective applications.

Blended learning is a dominant instructional model in business education, offering flexibility and accessibility while posing challenges around student self-regulation, engagement, and critical thinking. Artificial Intelligence (AI) integration within virtual learning environments offers opportunities to address these issues by personalizing support, scaffolding learning, and enhancing interaction. However, without thoughtful design, AI risks promoting automation bias, surface learning, and reduced critical thinking. This chapter proposes a framework for integrating AI into blended learning, grounded on three pillars: purposeful AI tools' alignment with cognitive outcomes, integrated design across modalities, and scaffolding for self-regulation and collaboration. A case study from a Strategic Management course in a UK Business School illustrates how AI functionalities can support deeper learning and engagement without eroding rigor. The chapter offers a pathway for leveraging AI while preserving analytical and ethical reasoning in business education.

This paper presents a new Deep Q-Network (DQN) method to improve the efficiency of solving the Dial-a-Ride Problem (DARP). We focus on the solution quality of DARPs, aiming to enhance the Total Travel Cost. Our approach begins by generating initial solutions through an insertion heuristic, which are then represented within a Markov Decision Process (MDP). Based on the DQN's ability to learn from experience, the system refines its decisions over time. A local search is integrated through the DQN process improving upon the baseline solutions. The results of our experiments demonstrate that the DQN leads to more efficient vehicle routes by improving the overall system performance for DARPs.

This study investigates the effectiveness of forecasting energy commodity prices using Artificial Intelligence-based models that account for the transition to cleaner energy sources during periods of significant market instability, such as the COVID-19 pandemic and the Russia-Ukraine conflict. Employing the Nonlinear Auto-Regressive model with exogenous inputs (NARX) over a comprehensive daily dataset from 2006 to 2023, the results reveal that machine learning models incorporating global energy transition factors perform better than traditional ANN and XGBoost models. The findings also reveal that integrating nonlinear relationships and external factors such as policy changes, technological advancements, and geopolitical events outperform traditional forecasting methods. This approach captures the complex dynamics of energy markets during periods of instability. By providing a reliable forecasting framework, this study enhances the understanding of energy market behaviors amid global transitions and uncertainties, promoting more adaptive and sustainable approaches to energy management. JEL Classification : C53, Q41, Q54

Sterlet is a perishable product; therefore, its freshness monitoring and shelf-life evaluation are important. In this study, a series of analytical techniques named physicochemical, microbiological, sensory, colorimetric, and mid-infrared and fluorescence spectroscopies were applied on Sterlet (Acipenser ruthenus) samples during 18 days of storage at 4 °C. The water content increased from 72.8 g/100 g on day 1 to 77.81 g/100 on day 14. Regarding the peroxide value (PV), the initial value was 4.17 meq/kg of Sterlet on day 1, reaching a maximum on day 4 (4.9 meq/kg of Sterlet), and then it decreased gradually, attaining a value of 0.7 meq/kg of Sterlet on day 18. Generally, the thiobarbituric acid reactive substance (TBARS), total viable count (TVC) and psychrotrophic count (PTC) increased during the storage time and increased from 0.03 to 0.13 MDA eq./kg of Sterlet sample, 2.27 to 9.09 log10 CFU/g, and 2.18 to 9.15 log10 CFU/g, respectively, on day 1 and 18, respectively. The microbiological and sensory analyses indicated that Sterlet samples were acceptable for human consumption up to 7 days of storage at 4 °C. This result was confirmed by fluorescence measurements, since the principal component analysis (PCA) applied to the NADH and MIR spectra allowed for a clear differentiation between Sterlet samples aged 7 days or less from the others. This trend was confirmed by the factorial discriminant analysis (FDA) applied to the NADH and MIR spectra, since a correct classification with leave-one cross-validation of 94.44% was observed. In addition, the heatmap of the Pearson correlation coefficients showed high correlations between overall acceptability and microbiology parameters and the structural properties of Sterlet samples during storage, indicating that the modifications observed at the macroscopic level were related to those notedat the molecular scale.