In a study published in the Journal of King Saud University Computer and Information Sciences, researchers led by Qiuyang Zhang and Hongfeng Chai introduced a weakly supervised machine learning framework that aligns decision-tree rules with graph representations through optimal transport to detect illicit anomalies across complex transaction networks. Detecting suspicious activities such as cryptocurrency money laundering or banking fraud has long presented severe difficulties because verifying ground-truth labels requires exhaustive and expensive human investigations. While previous semi-supervised techniques attempted to overcome label scarcity by generating flat pseudo-labels from heuristic rules or classifiers, they consistently overlooked relational structures and suffered from severe noise propagation. The newly developed architecture addresses this limitation by transforming hierarchical decision paths into a dedicated rule graph, projecting the resulting rule groups into class prototypes, and using optimal transport to smoothly couple those prototypes with continuous transaction embeddings. On the benchmark Elliptic Bitcoin transaction graph, this integrated methodology attained an area under the precision-recall curve of 0.687, outperforming leading graph neural baselines.
Financial crime detection has evolved rapidly as decentralized blockchain architectures and complex electronic banking systems expanded globally. While public cryptocurrency ledgers record transfers chronologically within blocks linked by cryptographic hashes, pseudonymous account structures offer avenues for illicit capital flight, scam operations, and cybercrime. Early automated screening systems in traditional banking predominantly relied on human-crafted heuristics and if-then logic focused on transaction amounts, velocity, or high-risk geographic flags. Although transparent and computationally lightweight, static rule-based systems generate excessive false alarms and fail to keep pace with sophisticated evasion schemes. Conversely, modern graph neural networks such as Graph Convolutional Networks and GraphSAGE learn directly from raw network connectivity to identify mixer interactions and laundering paths. However, supervised graph architectures require massive quantities of verified ground-truth annotations, which financial institutions rarely possess. Existing weakly supervised methods that translate heuristic rules into independent hard labels discard the relational context of complete decision sequences, creating a severe semantic mismatch between discrete logical rules and continuous feature representations.
To reconcile discrete heuristic decision logic with continuous graph learning, the authors designed a dual-branch framework that structures rule knowledge before transferring it to unlabeled transactions. In the rule branch, the system trains an adaptive-depth decision tree exclusively on available labeled training samples using the Gini impurity criterion. Rather than deploying the tree purely as an isolated classifier that assigns flat pseudo-labels, the researchers map all transactions through root-to-leaf traversals to identify their terminal leaves. Transactions that follow the identical complete path of feature splitting conditions are grouped together to construct a specialized rule graph. Within this formulation, decision-tree nodes themselves do not serve as network vertices; instead, individual transactions constitute the graph nodes. An undirected edge connects two distinct transactions if and only if both share the exact same complete root-to-leaf decision sequence. Subgraphs sharing only partial prefix splits remain completely disconnected. This grouping strategy preserves structural path-level dependencies that traditional flat pseudo-labeling pipelines systematically discard.
Once complete-path groups are formed, the rule branch refines their representations using graph convolution. Because every complete-path group constitutes an isolated clique with self-loops, normalized graph convolutional propagation algebraically reduces to group-mean aggregation followed by a learnable linear transformation and activation function. This message passing operation condenses transactions governed by identical decision criteria into unified representations, which are subsequently aggregated into two primary class prototypes representing licit and illicit behavior. Simultaneously, the data branch operates entirely in parallel, encoding raw local transaction attributes and neighborhood-based structural vectors into dense continuous vectors through a Multi-Layer Perceptron. Crucially, the data branch processes feature vectors without message passing across original transaction edges, keeping the two representation spaces fundamentally distinct. By maintaining separate rule and data branches, the framework preserves interpretable structural heuristics on one side while extracting flexible continuous data representations on the other, setting the stage for probabilistic cross-branch alignment.
Bridging the semantic gap between continuous data-driven representations and discrete rule prototypes represents the core architectural milestone of the proposed system. Rather than relying on simple vector concatenation, average pooling, or maximum mean discrepancy, the researchers adopted an optimal transport framework based on the Kantorovich formulation. Optimal transport computes a globally constrained transport plan that minimizes the total cost of mapping the empirical distribution of continuous transaction embeddings onto the discrete distribution of rule prototypes. The model calculates cost matrices from normalized Euclidean distances and stabilizes the optimization by incorporating an entropy regularization term solved via iterative Sinkhorn algorithms. By enforcing uniform marginal constraints across all transaction nodes and the two prototype states, the resulting transport plan constructs an admissible joint coupling. Normalizing each row of this optimal transport plan generates a valid conditional probability distribution, providing every unlabeled transaction with an explicit, confidence-weighted soft pseudo-label that directly supervises downstream classification via a cross-entropy objective.
While soft optimal transport assignments provide continuous probabilistic guidance, noisy rule transfers can still degrade classifier boundaries under severe class imbalances. To prevent ambiguous correspondences from misguiding the learning process, the framework incorporates cross-branch contrastive learning governed by a dynamic median-based filtering mechanism. At every training update, the system computes the median assignment confidence across all unlabeled nodes and scales it by a validation-tuned multiplier to establish an adaptive retention threshold. Only unlabeled transactions whose maximum prototype assignment exceeds this threshold enter the contrastive objective, while less confident samples remain restricted to soft supervision. For every retained sample, the contrastive loss minimizes embedding distance toward the assigned positive rule prototype while simultaneously maximizing separation against the opposing class prototype. By transforming confident optimal transport couplings into metric constraints, the contrastive module enhances intra-class compactness and sharpens decision boundaries, actively counteracting the noise propagation that frequently undermines weakly supervised graph learning.
The investigative team verified their architecture across four benchmark graph datasets featuring varied structural connectivity, feature semantics, and label distributions. On the Elliptic Bitcoin dataset, which comprises 203,769 nodes and 234,355 directed edges spanning 49 distinct time intervals with 9.8 percent illicit transactions, the full model achieved superior performance. It recorded an area under the receiver operating characteristic curve of 0.894, an area under the precision-recall curve of 0.687, and a top-ten-percent F1-score of 0.561. This marked a notable improvement over the strongest competing baseline, BWGNN, which registered 0.551 in precision-recall area. On the Amazon e-commerce review graph, containing 11,944 user nodes and more than 4.39 million edges modeling suspicious reviewing relationships, the model attained an area under the precision-recall curve of 0.905 and tied the highest top-ten-percent F1-score of 0.865 alongside LSTM and CPL baselines, confirming strong ranking capability.
Evaluating the framework across the remaining datasets illuminated key boundary conditions and nuances of rule-to-data knowledge transfer. On the Reddit user interaction dataset, which contains 10,984 nodes and 168,016 edges identifying banned users, the proposed model led all baselines but attained an area under the precision-recall curve of only 0.090. The authors noted that underlying text embeddings and comment topologies provided weak intrinsic class separability, restricting the discriminative power of distilled decision-tree rules. On the real-world Libra Bank dataset, containing 385,100 accounts and 597,165 transfer edges with an extreme fraud ratio of just 0.16 percent, standard feature-based models like support vector machines scored perfect precision-recall areas because input features included direct alert and report counts that defined the ground-truth labels. When researchers implemented a leakage-controlled protocol removing those four shortcut features, the proposed framework raised its precision-recall area from 0.0954 to 0.1643, outperforming specialized anomaly detector FraudGT by 0.0999.
Comprehensive ablation experiments and diagnostic evaluations confirmed that every individual framework component plays a decisive role in stabilizing training and boosting anomaly detection. On the Elliptic dataset, removing the rule graph entirely caused the precision-recall area to plunge from 0.674 down to 0.456, while eliminating the optimal transport loss dropped performance to 0.295 with substantial run-to-run variance. Stripping away contrastive learning reduced precision-recall area to 0.272. In post-hoc pseudo-label quality diagnostics on Elliptic, introducing contrastive learning elevated pseudo-label precision from 0.0355 to 0.1186 and pseudo-label F1-score from 0.0631 to 0.1759, primarily by eliminating false-positive over-predictions. However, on leakage-controlled Libra Bank, contrastive learning slightly reduced pseudo-label precision from 0.2516 to 0.2429, proving that contrastive refinement functions as a dataset-dependent regularizer rather than a universal denoiser. In computational efficiency tests using an NVIDIA GeForce RTX 4070 Laptop GPU, the model achieved an inference latency of 8.14 milliseconds, substantially outperforming BWGNN and FraudGT.
Despite its empirical advantages, the researchers highlighted several technical constraints and methodological limitations that delineate the scope of their findings. The framework operates strictly within individual datasets to achieve cross-branch representation alignment rather than cross-institutional domain adaptation. Errors originating within the initial decision tree can propagate directly into rule graphs, prototypes, and pseudo-labels if training labels are unrepresentative. Furthermore, assuming uniform marginals across rule prototypes becomes restrictive when evaluating extreme class imbalances, as observed in banking networks. Although transporting two class prototypes exhibits linear computational complexity relative to node count, repeatedly constructing cost matrices and executing Sinkhorn iterations introduces noticeable overhead on massive graphs. The data branch currently lacks original graph convolution and temporal modeling. Future work will investigate ensemble decision paths, low-rank or mini-batch optimal transport solvers, and multi-modal integration to enhance scalability and robustness across dynamically evolving transaction environments.
Subject of Research: Weakly supervised anomaly detection on transaction graphs using optimal transport and contrastive learning
Article Title: Rule-graph alignment via optimal transport for weakly supervised anomaly detection on transaction graphs
Article References: Zhang, Q., Chen, K., Jiang, M., Shui, Y., Tan, Y., Li, Z., Zhu, H., & Chai, H. (2026). Rule-graph alignment via optimal transport for weakly supervised anomaly detection on transaction graphs. Journal of King Saud University Computer and Information Sciences, 38(9), Article 976. https://doi.org/10.1007/s44443-026-01381-3
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01381-3
Keywords: anomaly detection, optimal transport, transaction graphs, graph neural networks, decision trees, anti-money laundering, weakly supervised learning, contrastive learning, Elliptic dataset, cryptocurrency forensics, Bitcoin transactions, pseudo-labeling, Sinkhorn algorithm, Kantorovich formulation, graph convolutional networks, financial crime detection, Libra Bank dataset, class imbalance
Cite Scienmag News
Denise Maddox. (November 21, 2026). Rule Graph Alignment via Optimal Transport Improves Weakly Supervised Anomaly Detection. Scienmag. https://scienmag.com/rule-graph-alignment-via-optimal-transport-improves-weakly-supervised-anomaly-detection/
Denise Maddox. "Rule Graph Alignment via Optimal Transport Improves Weakly Supervised Anomaly Detection." Scienmag, 21 November 2026, https://scienmag.com/rule-graph-alignment-via-optimal-transport-improves-weakly-supervised-anomaly-detection/. Accessed 21 November 2026.
Denise Maddox. "Rule Graph Alignment via Optimal Transport Improves Weakly Supervised Anomaly Detection." Scienmag. November 21, 2026. https://scienmag.com/rule-graph-alignment-via-optimal-transport-improves-weakly-supervised-anomaly-detection/








