Conference Agenda

Deals among organized crime groups (OCGs) can turn sour and may lead to conflict – sometimes violent. Yet, conflicts can be set aside when a new, maybe lucrative, joint opportunity arises. Building on previous works on co-offending and network violence among OCGs, we bring together cooperation and conflict and study their co-evolution over time. For this, we leverage police record data (2004-2015) provided by Thames Valley Police – the largest non-metropolitan police force in England and Wales. The data consist of 25,977 organized crime-related events in which at least one OCG member was involved. In the study, we focus on 147 OCGs, who were involved in 79 cooperation and 141 conflict events.

We study these events using relational event analysis through dynamics network actor models (DyNAMs). These models were proposed for relational event analysis from an actor-oriented perspective. To be able to study the rate of cooperation in this modeling framework, we propose new methodology to estimate rate models for undirected relational events using ideas from multiple imputation. Having estimated both rate models and partner choice models for cooperation and conflict, we find a strong interdependence between the two criminal networks among the OCGs. Yet, the event rate and partner choice of the groups are differentially affected by their network position (e.g., degree, embeddedness) and partner characteristics (e.g., primary criminal activity) for the two types of relational events. As the co-evolution of cooperation and conflict in the criminal context so far remained largely unexplored, this study fills a gap in our understanding of the dynamics underpinning organized crime operations.

Aims

This study explores the duality between criminal actors and relational hyperevents in an Italian mafia network, focusing on co-offending and communication. Relational hyperevents, defined as time-stamped interactions between two or more actors, provide a granular view of network evolution.

We distinguish two ontologically different event types: communication (sharing information/resources to commit a crime) and co-offending (the crime’s realization). We argue that these event types are mutually constitutive—communication precedes co-offending in organizing criminal activity. This reasoning extends beyond criminal networks to scientific, political, and artistic networks, where collaborative interactions (e.g., communication) precede outcomes (e.g., publications, legislation, or artistic productions).

To test this, we analyze three multiplex relational mechanisms—triadic closure, interaction repetition, and interaction accumulation—alongside their uniplex counterparts and exogenous factors (gender, age, leadership, locale affiliation, and kinship). Gender, age, and leadership are examined through selection/activity and homophily, while kinship and locale affiliation are assessed in relation to communication and co-offending.

Data & Methods

We use court records from Italian mafia investigation Infinito. Infinito documents the 'Ndrangheta’s drug-trafficking and territorial expansion in Lombardy, comprising 353 actors and several hundred events.

We apply Relational Hyper-Event Models (RHEM), which conceptualize polyadic interactions as hyperedges in a hypergraph. RHEM, combined with high-temporal-resolution data, allows us to model the interplay between co-offending and communication. Using the eventnet software package, we operationalize relational mechanisms and generate alternative (unobserved) non-events. We then estimate effects via a Cox proportional hazards model, identifying what characteristics make future events more or less likely.

Results

Preliminary results indicate that the network is held together primarily by communication ties, while co-offending is fragmented, sometimes occurring as single-actor crimes. Ignoring the relational hyperevent nature of the data could lead to significant information loss. Communication events (meetings and phone calls) are moderately correlated, and co-offending lags behind communication events.

Our RHEM analysis supports our hypothesis: communication precedes co-offending, especially in interaction repetition (within actor subsets). Notably, we find no evidence of triadic closure, contradicting existing social and criminal network research. Instead, interaction repetition explains observed closure effects. Additionally, uniplex interaction repetition and age- and leadership-based activity and homophily are significant for both event types.

Conclusions

We demonstrate RHEM’s utility in criminal network research and propose that traditional triadic closure effects may be artifacts of data aggregation rather than genuine actor tendencies. This has implications for criminal network studies using archival data.

Our findings bridge the gap between literature on the social organization of crime and co-offending networks, with potential applications in law enforcement, particularly in designing time-sensitive intervention strategies. Lastly, we suggest that distinguishing between process-oriented and outcome-oriented events can enhance network research in various domains, such as scientific collaboration and co-authorship analysis.

Urbanization has profoundly reshaped population distribution, with over half of the global population now living in cities, and has created new challenges in understanding urban issues such as crime. Despite this urban shift, the interplay between population dynamics, mobility, and crime remains poorly understood. In this work, we investigate how commuter flows, as a proxy for human mobility, can quantify the attractiveness of cities and shed light on differences in crime rates across urban regions in Great Britain.

We use commuting pattern data to explore criminological theories connecting population dynamics with crime [1]. While cities attract more people, they also experience higher levels of social anonymity, reduced community cohesion, and increased opportunities for crime [2, 3]. We employ journey-to-work census data for over 8 million workers in the UK to construct a directed network where nodes represent cities and edges capture commuter flows. We hypothesize that city attractiveness, as measured through the network, explains the non-linearity in how crime scales with population. We perform probabilistic scaling analysis [4] to assess the non-linearity in the relationships between attractiveness, commuter numbers, and crime types.

Our results reveal a positive correlation between commuter attractiveness and crime rates, with significant variations across crime types. Hub cities with high commuter inflows exhibit higher crime rates, supporting theories that transient populations weaken social control. Interestingly, certain crimes, such as theft, display superlinear scaling with attractiveness, while others, like burglary, align more closely with static population metrics.

This work advances the literature on urban crime by combining mobility network data and scaling analysis with criminological theories. By quantifying city attractiveness and examining its influence on crime, we present a novel framework for exploring the mechanisms underlying urbanization.

[1] Marcos Oliveira. “More crime in cities? On the scaling laws of crime and the inadequacy of per capita rankings—a cross-country study”. In: Crime Science 10.1 (2021), pp. 1–13.

[2] Claude S Fischer. “Toward a subcultural theory of urbanism”. In: American journal of Sociology 80.6 (1975), pp. 1319–1341. [3] Louis Wirth. “Urbanism as a Way of Life”. In: American journal of sociology 44.1 (1938), pp. 1–24.

[4] Jorge Leitao et al. “Is this scaling nonlinear?” In: Royal Society open science 3.7 (2016), p. 150649.

Crime is a complex phenomenon influenced by social, spatial, and contextual factors. Traditional statistical methods have long been used to analyse crime data, yet network-based approaches remain underexplored despite the inherently relational nature of criminal activity. Representing crime data as a network enables a more comprehensive understanding of its structure, capturing interactions between crimes, locations, and offenders that traditional methods often overlook.

This study examines a mid-size British police force dataset containing offender-victim relationships from crimes committed over a 46-months period. Given the dataset's richness, two network representations are constructed: a co-offending network, where a link between two suspects is established if they participated in the same crime event, and a directed network that links offenders to victims. To uncover latent structures within the data, we apply latent variable models that help provide deeper insights into broader crime patterns.

Specifically, we investigate how exposure to violence influences the behaviour of non‐violent offenders by assuming that an individual’s propensity for violent acts is influenced by the violent behaviours of connected actors and their position within a latent social space. Moreover, recognising that offender connections often arise from shared social, spatial, or contextual factors, we also cluster network edges to reveal underlying structures. This approach would allow to better understand the latent environments in which criminal associations form and explains why offenders participate in the same events. Ultimately, these methods offer valuable insights into the mechanisms driving criminal behaviours, providing a clearer, more nuanced perspective on UK crime patterns.

Purpose: In 2021, the US Marshals Service (USMS) partnered with law enforcement agencies across Galveston County, TX to conduct Operation Washout (OW), a 10-day operation intended to reduce violent crime through the execution of arrest warrants targeting gang members, firearm, and drug law violators. This study examines the impact of OW on the co-offending relationships of OW arrestees.

Methods: Arrest data for all of Galveston County are scraped from Police2Citizen Daily Bulletins and used to construct pre- and post-OW co-offending networks. These networks consist of individuals arrested during the intervention period, as well as all co-offenders within the first two neighborhoods. Stochastic actor-oriented models are estimated to model the rate that new co-offending relationships form following OW.

Results: Fugitives arrested as part of OW occupied brokerage positions in the network. Arrest during the intervention precedes a reduction in the rate that new co-offending relationships are formed. However, the rate that fugitives established new co-offending relationships was largely resistant to arrest.

Conclusions: Consistent with research on Operation Triple Beam, USMS-led fugitive operations appear to resemble broker targeting strategies proposed by network criminologists. Despite the selection of appropriate operational targets, fugitives’ ability to forge new co-offending relationships remains unimpeded by OW arrest.

Corruption, collusion, and fraud in public procurement (PP) have a long history in research (Fazekas et al., 2016; Dávid-Barrett and Fazekas, 2020); however, only some studies have accounted for the relational aspects of these phenomena—specifically, the network created by government agencies and suppliers linked through the common signature of a contract (Wachs and Kertész, 2019; Lyra et al., 2022).

Our study continues the analysis of PP as a network phenomenon, using national-level datasets released by the Mexican government over twelve years (2011–2022). We developed a method based on corruption indicators, network information, and sanction data to detect likely fraudulent contracts. We investigated the application of Positive-Unlabeled Learning (Jaskie and Spanias, 2022), a specific machine learning task that handles cases with only a subset of positive examples. We integrated different estimations of the unknown true proportion of the positive class (sanctioned suppliers) and evaluated their performance against a random model. Our methodology achieves an average precision score between 75% and 85% in the top 5% of prediction scores, compared to 30% for a random classifier.

Moreover, since we are interested in public policy interventions against fraud and corruption, we analyzed the feature importance of the model using Shapley Additive Explanations (SHAP) values (Lundberg and Lee, 2017). We discovered that the most influential features in the model are the clustering properties of the network. For example, the neighborhood proportion of direct contracts –contracts assigned with no competition–, the neighborhood average corruption risk index (Fazekas et al., 2016), and competitive clustering –the number of cycles over the number of paths of length three for a buyer (Wachs et al., 2021; Fazekas and Wachs, 2020)– are the three most important features of the model.

By analyzing the SHAP dependence plots of the model, we observed a non-linear relationship between SHAP values and the network features. For instance, in the neighborhood proportion of direct contracts, most likely positive (fraudulent) contracts are concentrated at middle-high levels of the feature, with a sharp decline in values close to one. This indicates that contracts in neighborhoods composed entirely of direct contracts are not necessarily the most fraudulent.

In conclusion, by combining multiple data sources and applying advanced machine learning techniques, we significantly improve the identification of high-risk contracts and suspicious players in the PP market, highlighting the relevance of network and contract characteristics.