TRACK: Multiple Instance Learning for Anti-Money Laundering in Transaction Networks
Organizations: KAIST, Daejeon, South Korea · Sogang University, Seoul, South Korea · LinqAlpha, New York, USA
Abstract
Anti-money laundering (AML) investigations are organized around accounts, while the suspicious evidence for an account may reside in only a few of its transactions. Yet many account-level detectors summarize transactions before learning, and message-passing models may aggregate once more over neighboring accounts. We name this discrepancy the aggregation mismatch and study it under an account-window benchmark label defined by the presence of at least one flagged transaction. We recast account-level AML detection as multiple instance learning, in which an account-window is a bag and its transactions are instances. We propose TRACK, which encodes every transaction with a shared encoder and selectively pools the encodings into an account decision. To emulate coarse supervision, we withhold the transaction labels of five AML datasets from TRACK throughout training. Under a time-ordered protocol, TRACK achieves the highest average precision among the tabular and graph baselines trained from account labels on every dataset. Despite using only account-window supervision, it remains competitive with edge classifiers trained directly from transaction labels. Its instance scores also rank flagged transactions well above chance and approach or exceed directly supervised ranking performance across datasets.
Figures & tables
| Group | Feature definitions | |
| Transfer | 6 | Log amount; direction; cyclic hour encoding; two channel indicators |
| Amount | 4 | Amount z-score and percentile rank; share of cumulative volume; frequency of similar prior amounts |
| Rhythm | 5 | Time since the previous transfer and its z-score; first-transaction indicator; prior transfer count; account tenure |
| Counterparty | 3 | New-counterparty indicator; prior transaction count and elapsed time for the account pair |
| Balance | 2 | Normalized outgoing–incoming differences in transaction count and amount |
| Pass-through | 2 | Time since the most recent incoming transfer; current-to-incoming amount ratio |
| Dataset | Accounts | Transactions | Account-window samples | Flagged accounts | Flagged transactions | Flagged test samples | Real data | Publicly available |
| AMLworld ( Altman et al., 2023 ) | 515,078 | 5,077,097 | 1,105,081 | 5,862 (1.14%) | 1,522 (0.03%) | 1,643 (0.46%) | ✗ | ✓ |
| AMLNet ( Huda et al., 2025 ) | 11,000 | 1,088,410 | 33,000 | 1,468 (13.35%) | 1,709 (0.16%) | 247 (2.25%) | ✗ | ✓ |
| AMLSim ( Suzumura and Kanezashi, 2021 ) | 9,999 | 1,323,234 | 29,996 | 1,639 (16.39%) | 1,719 (0.13%) | 423 (4.23%) | ✗ | ✓ |
| Elliptic++ ( Elmougy and Liu, 2023 ) | 265,354 | 1,661,916 | 268,650 | 14,266 (5.38%) | 15,863 (0.95%) | 1,747 (5.24%) | ✓ | ✓ |
| HOFINET | 452,816 | 4,732,130 | 758,994 | 9,644 (2.13%) | 14,190 (0.31%) | 2,723 (1.31%) | ✓ | ✗ |
| Method | AMLworld | AMLNet | AMLSim | Elliptic++ | HOFINET | |||||||||||||||
| AP | AUC | F1 | GM | AP | AUC | F1 | GM | AP | AUC | F1 | GM | AP | AUC | F1 | GM | AP | AUC | F1 | GM | |
| Tabular classification | ||||||||||||||||||||
| LR | 0.062 [-2pt] | 0.755 [-2pt] | 0.552 [-2pt] | 0.527 [-2pt] | 0.373 [-2pt] | 0.950 [-2pt] | 0.706 [-2pt] | 0.795 [-2pt] | 0.357 [-2pt] | 0.719 [-2pt] | 0.688 [-2pt] | 0.547 [-2pt] | 0.252 [-2pt] | 0.777 [-2pt] | 0.053 [-2pt] | 0.054 [-2pt] | 0.539 [-2pt] | 0.993 [-2pt] | 0.813 [-2pt] | 0.984 [-2pt] |
| MLP | 0.091 [-2pt] | 0.765 [-2pt] | 0.574 [-2pt] | 0.415 [-2pt] | 0.590 [-2pt] | 0.963 [-2pt] | 0.772 [-2pt] | 0.824 [-2pt] | 0.326 [-2pt] | 0.665 [-2pt] | 0.664 [-2pt] | 0.471 [-2pt] | 0.208 [-2pt] | 0.796 [-2pt] | 0.050 [-2pt] | 0.015 [-2pt] | 0.564 [-2pt] | 0.993 [-2pt] | 0.807 [-2pt] | 0.966 [-2pt] |
| XGBoost | 0.136 [-2pt] | 0.810 [-2pt] | 0.595 [-2pt] | 0.352 [-2pt] | 0.699 [-2pt] | 0.976 [-2pt] | 0.810 [-2pt] | 0.879 [-2pt] | 0.839 [-2pt] | 0.944 [-2pt] | 0.945 [-2pt] | 0.904 [-2pt] | 0.232 [-2pt] | 0.664 [-2pt] | 0.621 [-2pt] | 0.396 [-2pt] | 0.648 [-2pt] | 0.995 [-2pt] | 0.818 [-2pt] | 0.975 [-2pt] |
| General graph node classification | ||||||||||||||||||||
| Dataset | Recall@ | Recall@ | Recall@ | |||
| TRACK | Best | TRACK | Best | TRACK | Best | |
| AMLworld | 0.061 | 0.056 | 0.280 | 0.171 | 0.383 | 0.229 |
| AMLNet | 0.394 | 0.374 | 0.972 | 0.923 | 0.984 | 0.960 |
| AMLSim | 0.236 | 0.236 | 0.998 | 0.823 | 0.999 | 0.851 |
| Elliptic++ | 0.056 | 0.057 | 0.193 | 0.165 | 0.271 | 0.231 |
| HOFINET | 0.035 | 0.035 | 0.150 | 0.161 | 0.271 | 0.276 |
| Dataset | Method | Hit@1 | Hit@5 | MAP |
| AMLworld | Random | 0.329 | 0.765 | 0.423 |
| TRACK | 0.762 | 0.943 | 0.809 | |
| Multi-GIN | 0.791 | 0.964 | 0.818 | |
| FraudGT | 0.341 | 0.761 | 0.425 | |
| AMLNet | Random | 0.053 | 0.235 | 0.133 |
| TRACK | 0.976 | 0.977 | 0.931 |
| Variant | Computation and interpretation |
| Pre-aggregation | Raw transaction features are averaged before encoding. |
| MIL mean | Encoded transactions contribute uniformly. |
| Scalar max | One complete transaction score determines the account logit. |
| TRACK | Different transactions may supply different hidden dimensions. |
| Dataset | MIL mean | Attention | Scalar max | TRACK |
| AMLworld | 0.137 | 0.467 | 0.482 | 0.467 |
| AMLNet | 0.820 | 0.930 | 0.529 | 0.929 |
| AMLSim | 0.989 | 0.996 | 0.999 | 0.998 |
| Elliptic++ | 0.358 | 0.364 | 0.234 | 0.433 |
| HOFINET | 0.443 | 0.661 | 0.693 | 0.672 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Setting | Value |
| Learning rate | , |
| Hidden width | , |
| Encoder | 2-layer perceptron, ReLU, dropout |
| Optimizer | Adam, weight decay |
| Epochs | , early stopping at patience |
| Batch | bags |