Intégration des Techniques d'Apprentissage Automatique dans les Marchés Publics

Abstract

Given the increasing complexity of public procurement procedures and the need to enhance transparency and efficiency in offer selection, this research is based on a dataset extracted from the public procurement archives of the municipality of Mila and other sources. It aims to develop an automated predictive system capable of anticipating the winning candidate and the number of candidates, as well as the characteristics of the specifications, including technical offers. To achieve this, an approach combining machine learning, deep learning, and natural language processing (NLP) was adopted. Several high-performing algorithms such as Random Forest, Extra Trees, and KNN were compared after transforming textual data using TF-IDF and enhanced with data augmentation techniques. Deep models such as DNN and GRU were implemented to improve prediction accuracy, while retrieval-augmented generation (RAG) models enabled more intelligent contextual predictions. All these solutions were integrated into an interactive web interface allowing real-time predictions and efficient management of public procurement processes