[PhD defence] 02/09/2025 - Lucas Potin: "Analyse de graphes complexes pour détecter la corruption dans les marchés publics" (UPR LIA)

Mr Lucas POTIN will publicly defend his thesis entitled: "Analyse de graphes complexes pour détecter la corruption dans les marchés publics", directed by Mr Vincent LABATUT, Ms Christine LARGERON LETENO and Ms Rosa FIGUEIREDO, on Tuesday 2 September 2025.

Laboratory

• UPR 4128 LIA - Avignon Computing Laboratory

Composition of the jury

Mr Vincent LABATUT	LIA, Avignon University	Thesis supervisor
Ms Rosa FIGUEIREDO	LIA, Avignon University	Thesis co-director
Ms Christine LARGERON	Hubert Curien Laboratory, Jean Monnet University	Thesis co-director
Ms Peggy CELLIER	IRISA, INSA Rennes	Rapporteur
Mr Bruno CREMILLEUX	GREYC UMR 6072, University of Caen Normandy	Rapporteur
Ms Céline ROBARDET	LIRIS UMR 5205, INSA Lyon	Examiner
Pierre-Henri MORAND	JPEG, Avignon University / French Anti-Corruption Agency	Guest

Summary

Public procurement plays an essential role in the functioning of institutions, accounting for around 15 % of world GDP. In theory, the procedures are designed to guarantee transparency, competition and efficiency. In practice, however, they are often complex, difficult to understand, and exposed to risks such as collusion, favouritism and corruption. In this context, exploiting the large volumes of data available means that new ways of detecting fraud can be envisaged, in addition to traditional methods, particularly econometric ones. With this in mind, the DeCoMaP project (Detecting Corruption in Public Procurement), funded by the French National Research Agency, aimed to design detection tools combining legal, economic and IT expertise, based on data from French public procurement contracts. Carried out as part of the DeCoMaP project, this thesis targets two major methodological hurdles: the low reliability of existing databases, and the failure to take account of the relationships between economic players. In response, we adopt graph modelling to better represent the interactions between buyers and suppliers in public procurement. We begin by constructing two original databases, FOPPA and BeauAMP, based on official publications relating to French public procurement, distributed at national and European level. This work is based on in-depth data processing, including entity disambiguation. The resulting databases enable reliable relational analyses on a large scale, while surpassing existing sources in terms of quality and ease of use. They form a solid basis for the study of public procurement in France, and can benefit both researchers and public decision-makers.

From these graphs, we seek to distinguish the usual market networks from those with atypical, or even suspect, configurations. To do this, we extract discriminating patterns: sub-graphs that appear preferentially in one class rather than another. One of the major challenges lies in selecting the most relevant patterns for this task. To this end, we conducted a systematic study of 38 quality measures from the literature, comparing their behaviour and performance. Our results show that some measures are unstable, while others, which are simple but robust, perform well. We also introduce pattern clustering to limit redundancy and structure the pattern space more efficiently. Taken together, these results provide a useful benchmark for guiding the choice of quality measures in future work. Building on the previous results, we design the PANG framework (Pattern-based Anomaly detection in Graphs), which covers all the stages in the process: extraction, selection and representation of patterns, then graph classification. We are evaluating it both on standard public graph classification datasets and on the FOPPA database. The results show that PANG achieves comparable or even better performance than existing methods. Thanks to patterns that are easy to interpret, it makes it possible to give meaning to the configurations detected, and to offer both an economic and an institutional interpretation. This work contributes to better detection of irregularities, with the aim of strengthening the transparency and integrity of award procedures in public procurement.

Keywords : Graph theory, Machine learning, Public procurement