In online professional training, early identification of learners encountering difficulties is critical for enabling prompt, targeted pedagogical interventions. This study aims to develop a solution to predict at-risk learners at the earliest possible stage by leveraging comprehensive behavioral data extracted from a Learning Management System (LMS) over a seven-year period. We extract 12 quantitative behavioral indicators from detailed LMS logs, capturing diverse facets of learner engagement, such as time spent on activities, frequency of content interactions, and evaluation performance. By integrating these indicators, we develop a predictive framework that evaluates five different models. We conduct a misclassification analysis to identify and understand the sources of erroneous predictions. Based on these insights, we augment the feature set by introducing new indicators to better contextualize quiz-related interactions. Comprehensive hyperparameter optimization further refines the model parameters, balancing sensitivity and specificity effectively. Collectively, these steps significantly improve the model’s robustness, culminating in an optimized Random Forest classifier that achieves an F1 score of 82 %.

This paper presents a novel approach for real-time traffic accident prediction that leverages a clustering-based methodology and integrates multiple data sources, including historical accident records, traffic conditions, weather information, and road infrastructure characteristics. The proposed system identifies accident-prone zones by analyzing geospatial patterns in accident occurrences and clusters these zones based on multi-criteria similarities. Subsequently, six predictive models are trained separately on each cluster to capture context-specific risk factors. The performance of the cluster-based models is evaluated against a global model trained on the entire dataset, showing that specialized models outperform the global model in detecting accident occurrences, with significant improvements in recall and F1-score. Further, the generalization ability of the approach is assessed by predicting accident occurrences in previously unseen zones, demonstrating the effectiveness of the clustering approach. To enhance interpretability, the integration of LIME (Local Interpretable Model-agnostic Explanations) is proposed to provide textual explanations for each prediction, supporting informed decision-making and proactive accident prevention.

This paper introduces a solution designed to predict accident risk in the most accident-prone zones in France. Drawing on a comprehensive dataset of 210,007 accidents spanning four years, 22 high-risk zones are identified for focused models development. The predictive models leverage 35 features from four data sources, enabling a tailored approach to the identified zones. To enable real-time predictions, non-accident scenarios are generated based on temporal and spatial dimensions of each accident, resulting in a dataset comprising 770,880 observations. The dataset exhibits a significant imbalance, with only 8.2% representing accidents. To address this challenge, we employ the SMOTE technique. Then train and evaluate 11 predictive models. The initial version of the Weighted Classification Model (WCM) proposed in the literature is adapted to our context. Additionally, an enhanced version of the WCM is introduced, incorporating Exponential Ponderation to refine the weighted classification score. The Evaluation results demonstrate improved predictive performances, particularly in terms of precision, recall, and F1 score. The enhanced WCM exhibits a 1.3% improvement in the F1 score compared to the classical version. Real-time evaluations are conducted in France and showcase promising results, underscoring the practical applicability of the proposed solution.

Traffic accident prediction is an interesting research domain with the goal of providing analysis and methods to prevent road accidents. This paper presents a solution for predicting, in real-time and as early as possible, the occurrence and the severity of traffic accidents in France. For this purpose, a dataset of more than 200,000 observations collected over three years is used. We consider two accident classes, C1 for minor accidents and C2 for serious accidents; and a third class for non- accidents cases (C0). The dataset includes explanatory variables related to meteorological conditions, road traffic flow, and road network infrastructures. Initially, eleven predictive models are tested, and the results show discrepancies in predicting the observations of the three classes. To address this, a Weighted Classification Model (WCM) is proposed. WCM relies on the best models for each class (C0, C1 and C2) and penalizes them based on their classifications using a weighting coefficient. The evaluation results demonstrate that WCM improves the F1 score by 4% for C1 and 2.3% for C2, and achieves the best F1 macro score (the average F1 score over the three classes).

Road traffic safety is a major concern for road authorities and ordinary citizens. Consequently, accident prediction has become an interesting research topic that tries to provide solutions to predict, in real-time, traffic accidents occurrence and their severity. In this paper, a meta-model to predict, as early as possible, the risk and the severity of traffic accidents is proposed. The meta-model exploits seven predictive algorithms widely used in the literature and relies on a majority voting mechanism to improve predictions. For this purpose, a dataset of more than 45,000 observations is used, and two accident levels are considered. Based on the localization and the time of each accident, non-accident data are generated to create negative observations in the final dataset. Moreover, several features related to traffic flow, weather, and road conditions are collected and used as predictors to build and evaluate the predictive solutions and the meta-model. The experiment results show that the proposed meta-model dominates all other models in terms of F1 score.