Mafia infiltration in local government poses a severe threat to institutional integrity and economic development in Italy. Detection currently relies on the prefectural inspections provided for by Art. 143 TUEL (Testo Unico degli Enti Locali), which however affect only those Italian municipalities where a prior suspicion of infiltration already exists, managing each year to scrutinise only a handful of the nearly 8,000 Italian municipalities. A growing body of research has therefore begun in recent years to address the problem computationally, showing how organized crime infiltration in local administrations produces distortions that leave a detectable trace in budget accounts, procurement activity or local electoral dynamics. These data, after being converted in tailored indicators, can be exploited by classifiers to differentiate captured administrations from ordinary ones in the years preceding dissolution. Even so, existing literature has relied on cross-sectional or panel data, treating municipalities as isolated entities, disregarding both the inherent temporal component of the phenomenon and the procurement network through which local governments and corporate entities interact. Furthermore, their target definitions do not isolate the period in which infiltration actually unfolded: while some restrict it to the single year preceding dissolution, others label as infiltrated the entire mandate of the dissolved council, treating however the period of extraordinary commissioner-ship as ordinary management, even though the administrative record of those years reflects a corrective action carried out by state-appointed commissioners, rather than the conduct of an elected council. Addressing these gaps, this thesis formulates mafia detection in Italian municipalities as a supervised learning problem and investigates whether architectures whose inductive bias matches the structure of the underlying phenomenon can improve the identification of captured administrations. In particular, this work aims at isolating the predictive value obtained when transitioning from a standard static baseline assuming independent and identically distributed observations (XGBoost), to a sequential model capturing longitudinal dynamics (Long ShortTerm Memory - LSTM), and finally to a Temporal Graph Neural Network (Graph Convolutional LSTM), which jointly exploits relational and temporal structures. Hence, heterogeneous data sources are integrated and structured to match the inductive biases of each model. The framework combines 40 municipal financial indicators from MEF’s OpenBDAP portal, 49 engineered procurement features from ANAC’s BDNCP database, and the Rozes Score, a proprietary firm-level risk metric. Crucially, while static and sequential baselines observe each administration exclusively through its own administrative features, the proposed Temporal GNN explicitly translates public procurement into a dynamic heterogeneous graph, where municipal nodes are connected to corporate contractor nodes whenever a contract is awarded within a given year, following the idea of an inherently relational nature of organized crime. Finally, all architectures are benchmarked under a strict evaluation protocol to assess whether incorporating sequential dynamics and relational message passing enhances the identification of rare infiltration events, providing a data-driven predictive tool for anti-corruption and law enforcement authorities.

From Data Integration to Graph-Based Detection of Mafia Infiltration in Italian Municipalities: A Feature Engineering Approach on Public Procurement Data and Plan of Indicators

COVELLI, ALICE
2025/2026

Abstract

Mafia infiltration in local government poses a severe threat to institutional integrity and economic development in Italy. Detection currently relies on the prefectural inspections provided for by Art. 143 TUEL (Testo Unico degli Enti Locali), which however affect only those Italian municipalities where a prior suspicion of infiltration already exists, managing each year to scrutinise only a handful of the nearly 8,000 Italian municipalities. A growing body of research has therefore begun in recent years to address the problem computationally, showing how organized crime infiltration in local administrations produces distortions that leave a detectable trace in budget accounts, procurement activity or local electoral dynamics. These data, after being converted in tailored indicators, can be exploited by classifiers to differentiate captured administrations from ordinary ones in the years preceding dissolution. Even so, existing literature has relied on cross-sectional or panel data, treating municipalities as isolated entities, disregarding both the inherent temporal component of the phenomenon and the procurement network through which local governments and corporate entities interact. Furthermore, their target definitions do not isolate the period in which infiltration actually unfolded: while some restrict it to the single year preceding dissolution, others label as infiltrated the entire mandate of the dissolved council, treating however the period of extraordinary commissioner-ship as ordinary management, even though the administrative record of those years reflects a corrective action carried out by state-appointed commissioners, rather than the conduct of an elected council. Addressing these gaps, this thesis formulates mafia detection in Italian municipalities as a supervised learning problem and investigates whether architectures whose inductive bias matches the structure of the underlying phenomenon can improve the identification of captured administrations. In particular, this work aims at isolating the predictive value obtained when transitioning from a standard static baseline assuming independent and identically distributed observations (XGBoost), to a sequential model capturing longitudinal dynamics (Long ShortTerm Memory - LSTM), and finally to a Temporal Graph Neural Network (Graph Convolutional LSTM), which jointly exploits relational and temporal structures. Hence, heterogeneous data sources are integrated and structured to match the inductive biases of each model. The framework combines 40 municipal financial indicators from MEF’s OpenBDAP portal, 49 engineered procurement features from ANAC’s BDNCP database, and the Rozes Score, a proprietary firm-level risk metric. Crucially, while static and sequential baselines observe each administration exclusively through its own administrative features, the proposed Temporal GNN explicitly translates public procurement into a dynamic heterogeneous graph, where municipal nodes are connected to corporate contractor nodes whenever a contract is awarded within a given year, following the idea of an inherently relational nature of organized crime. Finally, all architectures are benchmarked under a strict evaluation protocol to assess whether incorporating sequential dynamics and relational message passing enhances the identification of rare infiltration events, providing a data-driven predictive tool for anti-corruption and law enforcement authorities.
2025
From Data Integration to Graph-Based Detection of Mafia Infiltration in Italian Municipalities: A Feature Engineering Approach on Public Procurement Data and Plan of Indicators
Graph Neural Network
Mafia Infiltration
Feature Engineering
Procurement Data
Plan of Indicators
File in questo prodotto:
File Dimensione Formato  
Master_Thesis_Covelli_Alice.pdf

Accesso riservato

Dimensione 1.54 MB
Formato Adobe PDF
1.54 MB Adobe PDF

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/115829