A temporal graph network (TGN) algorithm is introduced to identify fraudulent activities within a digital platform. The central premise is that digital transactions can be modeled via a graph network where various entities interact. The data used to build an event-based temporal graph (ETG) were sourced from an online payment platform and include details such as users, cards, devices, bank accounts, and features related to all these entities. Based on these data, seven distinct graphs were created; the first three represent individual interaction events (card registration, device registration, and bank account registration), while the remaining four are combinations of these graphs (card–device, card–bank account, device–bank account, and card–device–bank account registration). This approach was adopted to determine if the graph’s structure influenced the detection of fraudulent transactions. The results demonstrate that integrating more interaction events into the graph enhances the metrics, meaning graphs containing more interaction events yield superior fraud detection results than those based on individual events. In addition, the data used in this work correspond to Latin American payment transactions, which is relevant in the context of fraud detection since this region has the highest fraud rate in the world, yet few studies have focused on this issue.

El problema de los fraudes cibernéticos ha ido en aumento y es ya una problemática económica para las empresas que usan pagos electrónicos. Se han propuesto modelos y algoritmos dentro del área del aprendizaje automático con la finalidad de detectar patrones en las transacciones digitales que pudiesen exhibir las transacciones fraudulentas. Explicamos aquí una propuesta del uso de redes neuronales que usan estructuras de grafos para modelar y realizar la clasificación de usuarios fraudulentos.

The development of new objective functions in Neural Architecture Search (NAS) may include temporal and dynamic aspects of the problem to be modeled. In this line, we show the application of a temporal graph neural algorithm for detecting fraud on real data of an online payment platform, which is made up of different events that a user can perform, such as card registration, device registration, bank account registration, and IP registration. A combination of previous information generates different Event-Based Temporal Graphs (ETG) used in constructing the objective function for our neural network. The results show that combining different events effectively impacts AUC and recall, which is directly related to the proportion of fraudulent observations captured, and that there is a direct correlation between better metrics values and the graph density (measured as the proportion between vertices and edges). Few works focus on applying ETGs for fraud detection using real data and its analysis through the characteristics of the different ETGs, as we do in this work. Furthermore, our application works on real data from Latin America, the region with the highest fraud rate worldwide.

El fraude transaccional en plataformas digitales ha incrementado en años recientes. Los defraudadores utilizan diferentes técnicas para apropiarse de los recursos económicos de usuarios y sacar provecho de algunas vulnerabilidades. Por esta razón, se han implementado diferentes soluciones basadas en el aprendizaje automático para abordar el problema. En este trabajo se describen algunas características del fraude digital así como los principales métodos que se han utilizado para la detección de fraude. A su vez, la transición que ha existido hacia un enfoque que combina grafos con aprendizaje automático, ha propuesto las llamadas redes neuronales de grafos. Finalmente, se mencionan algunos retos del área relacionados a las características de los datos necesarios para operar este tipo de algoritmos.

Digital fraud poses a critical global challenge, particularly in Latin America, requiring detection systems capable of identifying illicit patterns in highly imbalanced transactional data. Recent work on graph neural networks for temporal heterogeneous data has explored several architectural directions, including type-aware aggregation with universal memory, relation-aware convolutions, memory-free continuous-time methods, and snapshot-based hierarchical state updates. Across these approaches, the persistent state mechanism is either shared across all entity types, absent entirely, or operates between snapshots rather than between events. We identify per-type, event-driven memory as a previously unexplored architectural axis and propose HetMem-TGN (Heterogeneous-Memory Temporal Graph Network), a continuous-time event-based architecture that maintains separate persistent memory modules for each vertex type and each edge type, with its own dimensionality and update dynamics. This allows the model to maintain distinct evolutionary trajectories for entities playing fundamentally different roles in fraudulent activity, rather than collapsing them into a shared memory space. We evaluate HetMem-TGN on real transaction data from a Latin American online payment platform, comprising 2.6 million events across six vertex types and five edge types, with severe class imbalance (0.33\% fraud rate). Across various subsampled distributions, HetMem-TGN consistently maintains high Precision and PR-AUC scores, substantially outperforming static heterogeneous baselines (HGT, HAN, RGCN) and dynamic homogeneous baselines (TGN, JODIE). Our experiments also reveal an inverse relationship between connectivity density and recall, with moderate undersampling outperforming the full dataset, providing actionable guidance for practitioners deploying graph-based fraud detection.

Fraud detection, with its rings and shared-attribute groups, is an obvious target for topological deep learning, yet we are aware of no higher-order fraud benchmark and no neutral test of whether lifting beats strong feature-only and pairwise baselines under one imbalance-aware protocol. This paper is that benchmark and that test: seven real transaction, review, and financial graphs lifted to candidate-cell hypergraphs, plus a generator D(mu) in which the value of structure is known by construction. The answer separates datasets into three regimes. In the feature-dominated regime node features already carry the signal: a feature-only model matches or beats the best pairwise and higher-order models, naive message passing amplifies camouflage, and learnable liftings give no significant gain, whether their cells are selected, generated, or regularized. In the column-foldable regime a temporal signal is strongly predictive but folds into node-level feature columns, so a lifting gains nothing over that cheaper encoding; signal being present is not sufficient for lifting to pay. Only the structure-rich regime rewards lifting. There the structure carries minority signal that no node-level encoding reaches, and a label-free temporal-coherence lifting beats the temporal columns by +1.9 AP points on the Elliptic Bitcoin graph (paired t(7) of about 12.7), a margin that survives feature augmentation on Elliptic++. One decisive test, applied uniformly to six temporal crossovers over five networks, is what locates that positive rather than selecting it. The generator, the lifting model, and the evaluation harness will be released with a future version.