Paper proposes a compliance anomaly detection algorithm for financial transaction data, addressing the challenges of behavioral complexity, structural dependence, and strong temporal characteristics. The method constructs a dynamic trajectory graph with multi-type edges based on account interactions, providing a unified representation of structural relationships and temporal order in financial transactions. A graph neural network is employed to embed node relationships and extract high-order behavioral structure features. A gated recurrent neural network is then used to model the temporal dynamics of trajectory sequences and capture the evolution patterns of transaction behaviors. In the classification stage, a joint optimization strategy combines contrastive learning with an anomaly scoring function to enhance intra-class cohesion and inter-class separation, improving the model's ability to identify abnormal behaviors. To evaluate the effectiveness of the proposed method, experiments are conducted across various sensitivity dimensions, including time window length, trajectory length, node dropout rate, anomaly ratio, and class imbalance. The results show that the method achieves superior detection performance and strong adaptability under diverse perturbation conditions, demonstrating the potential of structure-temporal joint modeling for financial compliance anomaly detection tasks. CCS CONCEPTS: Computing methodologies~Machine learning~Machine learning approaches.

This study addresses the growing complexity of transaction behaviors and the highly concealed nature of money laundering paths in current financial AML scenarios. It proposes a Transformer-based risk monitoring model for anti-money laundering. The approach is grounded in transaction sequence modeling and integrates the structural information of transaction graphs. A context-aware classifier is introduced to enable accurate identification and risk scoring of high-risk accounts. The model first applies feature embedding and positional encoding to each transaction. It then uses multiple Transformer layers to capture long-range behavioral dependencies. At the same time, it incorporates account interaction information from the graph structure. This enhances the model's ability to detect abnormal transaction chains across accounts. At the output stage, a classifier that fuses sequential semantics with graph context is used to determine the overall money laundering risk of each account. Multiple experiments were conducted on the publicly available Elliptic dataset. Results show that the proposed method outperforms existing mainstream models on evaluation metrics such as AUC, F1-Score, and Accuracy. It also demonstrates stronger discriminative power and greater stability in identifying high-risk accounts. Further analysis of model depth sensitivity and case-based verification supports the model's effectiveness in real-world complex transaction environments. The proposed method offers a more adaptable technical solution for financial institutions dealing with large-scale suspicious behavior detection tasks.