USDGNN: A Unified Spatiotemporal Dynamic Graph Neural Network Model for Credit Card Fraud Detection

Credit card fraud detection has evolved significantly over the past decade, as fraudulent activities have become more severe and continue to hinder the growth of digital payment environments. Researchers have conducted extensive studies on credit card fraud detection using traditional machine learning (ML) models, deep learning models, graph-based methods, and DGNN architectures. This section provides the most related studies, organized by methodological paradigm, and reveals the research gaps that motivate the current work.

2.1 Traditional machine learning and deep learning approaches

Early credit card fraud detection methods relied mainly on rule-based, statistical, and conventional machine learning techniques. Comparative data-mining studies identify severe class imbalance as a principal challenge affecting fraud detection performance [14]. Models such as LR, DT, SVM, XGBoost, and RF remain widely used because of their computational efficiency and interpretability; however, they generally treat transactions as independent tabular instances and depend strongly on predefined or handcrafted attributes. Ensemble methods such as RF and XGBoost have reported F1-scores of up to 0.89 on imbalanced fraud datasets, whereas classifiers such as SVM show weaker performance under severe class imbalance [15].

Data-level and algorithm-level imbalance-handling strategies, including SMOTE, Random Undersampling (RUS), and hybrid resampling techniques, have been extensively investigated for credit card fraud detection. Reported results indicate that combining SMOTE with RF can increase classification accuracy to 99.89%, while boosting-based methods such as CtRUBoost have achieved precision values of up to 95.7% [16]. Nevertheless, these approaches primarily address class-distribution imbalance and do not explicitly model structural relationships among transactions or their evolving temporal dependencies.

Hybrid deep-learning approaches have also combined autoencoder-based anomaly detection with GNN components and Lambda-style architectures for real-time fraud analysis; however, reported performance remained limited, with an accuracy of approximately 78% in one evaluated configuration [17].

2.2 Graph Neural Network for fraud detection

GNNs have emerged as an appealing paradigm for fraud detection because financial operations represented by transactions form a network in which users, merchants, devices, and cards are nodes and transactions are edges. GNN methods can capture structural anomalies that traditional ML methods fail to capture.

Early GNN-based fraud detection approaches commonly relied on static or homogeneous graph representations. Encoder–decoder graph architectures have represented consumers and merchants as graph entities and incorporated engineered attributes such as distance and rolling averages, achieving an F1-score of 0.86 and an ROC-AUC of 0.92 on the Sparkov synthetic dataset [18]. However, temporal information in such approaches is typically encoded as static attributes rather than through explicitly evolving graph states, limiting their ability to represent changes in fraudulent behavior over time. This hinders the model from adapting to changing fraud patterns.

Event-based temporal graph models provide an alternative to fixed snapshots by treating transactions as a continuous stream of timestamped events and using memory and gated recurrent unit (GRU)-based mechanisms to retain temporal context; reported performance includes an F1-score of 0.4523 and an AUC of 0.7255 [19]. More recent temporal graph approaches explicitly incorporate time-dependent information into graph learning. Causal temporal graph models combine causal inspection and intervention mechanisms with temporal graph features to improve generalization against evolving fraudulent behavior, with reported AUC and F1-score values of 0.825 and 0.721, respectively, on the FFSD dataset [20]. These approaches demonstrate the importance of temporal graph modeling but employ specialized architectures and evaluation protocols that differ from the snapshot-based USDGNN framework.

Recent reviews further demonstrate the increasing importance of graph representation learning for financial fraud detection. Systematic evidence indicates that GNN-based methods can exploit relational structures among financial entities and transactions that conventional tabular classifiers do not explicitly capture, while also identifying temporal and dynamic graph modeling as important directions for evolving fraud scenarios [21]. More recent analyses organize financial-fraud GNNs according to their graph-learning architectures and emphasize the importance of capturing complex relational patterns, architectural adaptability, and deployment considerations in real-world financial networks [22]. Graph-learning research has also highlighted scalability, robustness, interpretability, and the dynamic nature of financial networks as continuing challenges for practical fraud-detection systems [21]. These findings reinforce the motivation for USDGNN, which combines transaction-level graph construction with recurrent graph learning to model both structural and temporal information.

Our proposed credit card fraud detection system uses continuous time with a dynamic GNN framework and memory-based temporal modeling under realistic criteria of imbalanced and delayed feedback.

2.3 Research gap and motivation

Although machine learning, deep learning, and GNNs have significantly improved credit card fraud detection, several limitations remain. Conventional machine learning algorithms treat each transaction as an independent case and therefore do not capture relationships within the data. Although graphs improve dependency modeling, existing solutions mostly rely on static graphs and cannot capture time dependencies in financial transaction data. Temporal GNN-based solutions overcome this problem but lack a general framework for temporal graph generation. Therefore, a research gap exists in developing an integrated temporal graph learning framework that simultaneously captures structural relationships, temporal evolution, and severe class imbalance. To address this gap, this study proposes a USDGNN framework that combines chronological transaction ordering, KNN-based temporal graph snapshots, stacked GConvGRU learning, and cost-sensitive optimization into a unified pipeline for robust and reproducible credit card fraud detection. Table 1 highlights the methodological comparison of representative credit card fraud detection approaches and the proposed USDGNN framework.

Table 1 compares the methodological attributes of various representative credit card fraud detection models. Traditional machine learning algorithms fail to model structural or temporal relationships between transactions. Graph methods improve fraud detection by modeling interactions between transactions, but most rely on static graph representations or have limited ability to address class imbalance. Even though temporal graph algorithms take into account time relations, they apply specialized architectures for certain cases.

Table 1. Methodological comparison of representative credit card fraud detection approaches and the proposed Unified Spatiotemporal Dynamic Graph Neural Network (USDGNN) framework

Related cybersecurity studies further demonstrate the effectiveness of data-driven learning methods across different threat-detection tasks. Machine learning combined with feature selection has been applied to improve phishing website detection using URL-based features [21], while comparative evaluations of machine learning and deep learning have examined their effectiveness for network intrusion detection [22]. These studies further illustrate the broader applicability of learning-based approaches to cybersecurity detection problems.

Recent systematic reviews confirm that, despite substantial progress in GNN-based financial fraud detection, important challenges remain in modeling dynamically evolving transaction relationships, designing robust graph structures, and maintaining effective learning under highly imbalanced financial data [21-24]. These observations further motivate the proposed USDGNN framework, which integrates chronological transaction modeling, snapshot-based KNN graph construction, recurrent graph learning, and cost-sensitive optimization within a unified fraud-detection pipeline.

To ensure a leakage-free chronological evaluation, all reported experiments use the same 70–30 chronological partition defined in Section 3. We use transactions in the earlier 70% for model development, while retaining the later 30% as the held-out test set. We estimate feature-standardization parameters exclusively from the training partition and apply them unchanged to the test partition. Temporal batching and KNN graph construction are performed independently within each partition, ensuring that no preprocessing statistics, temporal snapshots, or graph edges cross the training–test boundary. The baseline classifiers are trained and evaluated using the same chronological training and test partitions to ensure a consistent comparison with USDGNN.

4.1 Testing environment

The experiment in this study assessed the efficacy of the USDGNN model framework for detecting fraud in credit card transactions within the constraints of recourse. All experiments were carried out on a personal computer to prove the practical feasibility of the research. Table 3 presents the full hardware and software specifications used in the experiment.

Table 3. Experimental testing environment

In this scenario, we not only provide a replicable environment but also demonstrate the usability of the suggested USDGNN framework for large-scale operations without requiring supercomputing hardware.

4.2 Results

We obtained all reported test results from the held-out final 30% of the chronologically ordered transaction stream. Model fitting, preprocessing-parameter estimation, and training-stage optimization were restricted to the earlier 70% training partition. The test partition was not used to estimate standardization parameters or optimize model parameters.

This section presents a detailed evaluation of experiments conducted with the proposed USDGNN framework for credit card fraud detection, highlighting the model's efficiency compared with traditional ML models (e.g., RF, LR, and NB). In this regard, the comparison evaluates the models' classification performance in the critical case of an unbalanced dataset and the need to capture and track dynamic transactions. In addition to numerical metric evaluations, the results include visualization metrics to evaluate classification efficiency, including ROC-AUC plots for both USDGNN and traditional ML models.

The proposed USDGNN-based credit card fraud detection model remains consistently stable during training. This stability is reflected in the loss evolution during training, where the loss decreases from 0.2854 at epoch 0 to 0.0067 at the last epoch (i.e., epoch 30). Table 4 shows the classification performance of the proposed temporal credit card fraud detection with USDGNN versus the classical baseline classifiers (e.g., RF, LR, and NB).

Table 4. Evaluation metrics comparison of Unified Spatiotemporal Dynamic Graph Neural Network (USDGNN) and baseline traditional machine learning (ML) models

All models show fairly good accuracy, but this is due to the natural class imbalance in the credit card fraud detection dataset; however, accuracy is not a dependable measure for fraud detection. In many similar cases, classification algorithms achieve high accuracy by favoring the majority target class and largely ignoring the minority target class. Thus, precision, recall, and F1-score measures are quite informative evaluation criteria that can be used to make fair comparisons.

Among the evaluated models, USDGNN achieved the highest F1-score of 0.8571, reflecting the strongest observed balance between precision and recall in the reported test split. Random Forest achieved the highest precision of 0.9189 but lower recall of 0.7727, whereas USDGNN achieved precision of 0.8478 and recall of 0.8667. Thus, the reported results do not indicate that USDGNN is uniformly superior across all metrics; rather, they show that USDGNN provides the highest F1-score and a comparatively balanced precision–recall profile among the evaluated classifiers.

The LR and NB models achieve extremely high recall; however, this high confidence comes at the cost of very low precision. Consequently, the F1-score recorded poor values. This limitation offers a crucial insight and highlights the limitations of linear and probabilistic algorithms on highly imbalanced datasets, where the model is biased toward predicting the majority target class. Our proposed USDGNN addresses this issue by constructing and exploiting graph neighborhood data, learning and modeling temporal correlations among transaction snapshots, and incorporating cost-sensitive learning throughout the optimization process. Consequently, within the reported test split, USDGNN achieved a comparatively balanced precision–recall profile, reflecting its ability to detect fraudulent transactions while limiting false-positive predictions relative to the high-recall baseline models. 

Figure 4 demonstrates the ROC curves of the proposed USDGNN and traditional baseline classification models. USDGNN achieved an ROC-AUC of 0.9913, indicating strong discrimination between fraudulent and legitimate transactions in the reported test set. However, RF and LR achieved higher overall AUC values. Therefore, the ROC analysis should be interpreted together with the precision, recall, and F1-score results rather than as evidence of uniform superiority of USDGNN across all evaluation criteria.

Figure 5 shows the precision-recall distribution of the evaluated models. This evaluation shows each model's effectiveness under an imbalanced class distribution. The RF model overruns the high-precision region, reflecting modest fraud detection that minimizes false alarms but abandons some detection capability.

4.png

Figure 4. ROC curves comparing the proposed Unified Spatiotemporal Dynamic Graph Neural Network (USDGNN) with baseline classifiers

5.png

Figure 5. Precision–Recall comparison of the evaluated classifiers

The LR and NB reside in the high-recall but low-precision region, demonstrating aggressive fraud-detection behavior that classifies many legitimate transactions as fraudulent. The proposed USDGNN occupies a comparatively balanced region of the precision–recall space, with a recall of 0.8667 and a precision of 0.8478. This position is consistent with its F1-score of 0.8571 and distinguishes its operating behavior from RF, which favors precision, and LR and NB, which favor recall at the expense of substantially lower precision. Table 5 lists selected studies on fraud detection, along with their techniques, data used, and performance results.

Since the studies summarized in Table 5 were conducted using different datasets, graph construction strategies, preprocessing procedures, and experimental protocols, the reported performance metrics are presented for contextual reference only and should not be interpreted as a direct empirical comparison with the proposed USDGNN framework. Instead, Table 5 illustrates the methodological evolution of fraud detection techniques from conventional machine learning methods to graph-based and temporal graph learning approaches, thereby motivating the development of the proposed USDGNN framework.

Table 5. Summary of representative credit card fraud detection studies and their reported performance

4.3 Sensitivity analysis of the K-Nearest Neighbors neighborhood size

To examine USDGNN's sensitivity to the KNN neighborhood size, we evaluated different values of k in the graph-generation process while keeping the temporal snapshot size fixed at 1,000 transactions. Accordingly, this analysis evaluates sensitivity to neighborhood connectivity only and should not be interpreted as a joint sensitivity analysis of all graph-construction parameters. The choice of neighborhood size directly controls the connectivity in the graph by determining how many transaction neighbors are connected to a single node within a particular temporal window; since information propagates throughout the graph during graph convolution via its topology, a suitable choice of k should provide balanced performance across all classes.

To examine the impact of k, the proposed USDGNN was trained and evaluated across neighborhood sizes of k = 5, 10, 15, 20, and 25, keeping all other parameters, including snapshot size, network architecture, optimizer, learning rate, and class weight settings identical across all configurations. It must be noted that, due to the absence of a separate validation set in the current experimental design, all reported sensitivity results were obtained by evaluating each k configuration directly on the same held-out chronological 30% test partition used for the main performance comparison in Section 4.2. Consequently, the selection of k = 15 as the reported best configuration was based on its observed test set F1-score performance rather than on an independent validation set. This represents a limitation of the current sensitivity analysis, the selected k value may be optimistically tuned to the specific test partition, and the reported performance differences across k values should therefore be interpreted as exploratory observations rather than as statistically validated conclusions. A rigorous future evaluation should employ a dedicated held-out validation partition, for example, a chronological 70/15/15 train, validation, test split, to select the optimal k independently of the final test set, followed by a single confirmatory evaluation on the held-out test partition. The sensitivity results are presented in Table 6.

From Table 6, we see that neighborhood size significantly affects the learning behavior of the proposed USDGNN. With a relatively small neighborhood (k = 5), the graph created is sparse, and information can’t easily propagate among neighborhood transactions. As a result, the model achieves relatively high accuracy but lower precision and F1-score.

Setting the neighborhood size to k = 10 significantly improves the graph's connectivity, thereby substantially improving both precision and F1-score. The maximum F1-score (0.8571) was observed at k = 15, accompanied by a reasonable precision (0.8478) and recall (0.8667), demonstrating a reasonable balance between sparsity and graph neighborhood connectivity in the USDGNN, allowing the capture of relevant structural connections without introducing excessive irrelevant information. Increasing the neighborhood size to k = 20 yields maximum recall (0.9111), whereby a larger fraction of fraudulent transactions is detected.

Table 6. Sensitivity analysis of the K-Nearest Neighbors (KNN) neighborhood size used for graph construction

However, this enhancement in recall comes at the cost of a significant drop in precision (0.6949) due to the addition of false-positive predictions from high information transmission between unrelated transactions, which builds denser graphs. Increasing the neighborhood size to k = 25 partially recovers precision, but the F1-score remains lower than at k = 15. Thus, this indicates that densely connected graphs are not always the solution for improving fraud detection.

Figure 6 shows the effect of the KNN neighborhood size on the performance of the proposed USDGNN. In this scenario, the USDGNN performs optimally at k = 15, as the balance between graph sparseness and neighborhood size is optimal; larger k yields a denser graph structure, degrading the classifier's performance, while lower k reduces propagation.

6.png

Figure 6. Effect of the K-Nearest Neighbors (KNN) neighborhood size on the performance of the proposed Unified Spatiotemporal Dynamic Graph Neural Network (USDGNN)

Although a larger graph allows more structural information to propagate, it also includes numerous edges between unrelated transactions, introducing additional noise. It is worth noting that there is a slight performance improvement when increasing k from 20 to 25; this suggests that USDGNN performs well over a relatively broad range of k; nonetheless, it does not outperform our configuration. The results indicate that USDGNN maintains competitive performance across the evaluated neighborhood settings, yielding the highest observed F1-score.

This conclusion is limited to variation in neighborhood size because the temporal snapshot size remained fixed at 1,000 transactions throughout the analysis.

4.4 Discussion

The experimental results indicate that the evaluated classifiers exhibit different performance trade-offs. USDGNN achieved the highest F1-score (0.8571) and accuracy (0.9991), with precision (0.8478) and recall (0.8667). In contrast, Random Forest achieved the highest precision of 0.9189 but a lower recall of 0.7727, whereas Logistic Regression and Naïve Bayes achieved higher recall values of 0.9773 but substantially lower precision. Accordingly, USDGNN should be interpreted as providing the strongest observed F1-based precision–recall balance among the evaluated models rather than as outperforming every baseline on every metric.

These results are also confirmed by the ROC analysis. USDGNN had an AUC of 0.9913 and an optimal operating point in the low false-positive region, which is critical for any fraud detection system to be practically usable. Furthermore, the training loss showed a gradual decrease from 0.2854 in the initial epochs to 0.0067 by the end of 30 epochs, supporting the stability and effectiveness of convergence. In addition, the analysis of sensitivity to neighborhood size indicated that the value k = 15 yielded the highest observed F1-score of 0.8571 compared to other values of k tested, i.e., k = 5, 10, 15, 20, and 25. Nevertheless, it must be pointed out explicitly that this choice is made using test set evaluation rather than using a separately reserved validation set since there is no validation set specified for the current experiment. Therefore, this conclusion is to be considered as sensitivity exploration rather than validation of the optimal configuration.

The outcomes are consistent with the most recent progress reported in the literature on graph-based fraud detection methods. Conventional machine learning methods suffer significant performance degradation under severe class imbalance because transactions are treated as independent tabular records without modeling structural or relational dependencies [14]. While ensemble-based approaches can enhance classification accuracy, they remain limited in their ability to capture structural relationships among transactions [15]. Furthermore, although resampling techniques such as SMOTE combined with ensemble classifiers can partially address class imbalance, they fail to capture the temporal dynamics of evolving fraudulent behavior [16].

Some of these limitations have been partially addressed through graph-based approaches. An Encoder-Decoder Graph Neural Network (EDGNN) architecture applied to the synthetic Sparkov dataset achieved an F1-score of approximately 0.86 and an AUC of 0.92 [18]. The proposed USDGNN obtained a comparable F1-score of 0.8571 when evaluated on the highly imbalanced real-world ULB credit card dataset, demonstrating that dynamic graph learning can be applied effectively to more challenging real transaction data. Among temporal graph learning approaches, the CaT-GNN framework, which incorporates causal and temporal graph modeling, reported an F1-score of 0.721 [20], while the Temporal Graph Network (TGN)-based approach for online payment fraud detection reported an F1-score of 0.4523 with an AUC of 0.7255 [19]. In contrast to these approaches, the proposed USDGNN provides a unified end-to-end framework that combines chronological transaction ordering, KNN-based temporal graph construction, stacked GConvGRU learning, and cost-sensitive optimization. However, these comparisons should be regarded as contextual validation only, not as direct empirical comparisons, because the referenced studies used different datasets, graph construction strategies, preprocessing techniques, evaluation methods, and fraud distributions.

Several observations can be drawn from the reported experiments. First, under the evaluated chronological split, temporal graph learning produced a favorable precision–recall trade-off relative to the traditional machine learning baselines, as reflected by the highest observed F1-score for USDGNN. USDGNN models structural relationships among similar transactions together with temporal relationships across sequential transaction snapshots, whereas the evaluated conventional classifiers operate on tabular transaction features without explicit graph-based temporal modeling. Additionally, USDGNN models structural relationships between similar transactions and temporal relationships between sequential transaction snapshots, whereas traditional classifiers only model the sequential transactions. This enables the framework to incorporate information from preceding transaction snapshots when classifying transactions in subsequent snapshots. Second, the severe class imbalance in the ULB dataset is addressed through cost-sensitive optimization rather than resampling. By assigning a larger loss contribution to misclassified minority fraudulent transactions, the training objective reduces the model's tendency to favor the majority legitimate class while preserving the original class distribution. Third, the neighborhood sensitivity analysis is consistent with this: it validates the importance of graph construction on the overall performance. The neighborhood size k = 15 is large enough to capture the structure of the graph yet small enough to avoid the “graph density” seen with larger values of k, maximizing the F1-score. Finally, the proposed framework is practical and scalable for large-scale fraud detection while keeping learning stable during training by combining KNN-based graph construction with recurrent graph convolution.

Although the proposed USDGNN framework showed promising results in terms of performance, the following limitations should be noted. The experiments used only the public ULB credit card fraud dataset, collected over a short period and anonymized with PCA-transformed features. Hence, the applicability of the proposed framework to other financial institutions, geographical regions, or new fraud techniques is an important issue that needs to be explored.

Second, the temporal snapshot size was fixed at 1,000 transactions throughout the experiments. Although a sensitivity analysis was conducted for the KNN neighborhood size and yielded the highest observed F1-score, an equivalent sensitivity analysis of the temporal snapshot size was not performed. Because snapshot size directly affects graph size, temporal granularity, and the amount of structural information available within each graph, the selected value of 1,000 should be regarded as a fixed design choice rather than an empirically optimized parameter. Future work should evaluate multiple snapshot sizes and jointly examine their interaction with the KNN neighborhood size. Third, only one chronological train–test split was examined, and repeated training runs, confidence-interval estimation, and statistical significance testing were not conducted. Consequently, the reported differences should be interpreted as observed performance differences under the present experimental setting rather than as statistically significant improvements. Future evaluations using multiple independent runs, additional datasets, confidence intervals, and appropriate hypothesis tests are required to establish the statistical reliability and generalizability of the observed performance differences.

Lastly, only three traditional machine learning baselines (random forest, logistic regression, and naïve Bayes) were included and evaluated comparatively. The USDGNN can be compared with other state-of-the-art GNN architectures, graph attention networks, and recent temporal graph learning models in future work using the same experimental setup and conditions. Moreover, incorporating real-time streaming transaction analysis, adaptive graph generation, explainable graph learning, and online continual learning would further enhance its use in real-world financial fraud detection systems.