Context-Guided Traversal with Deep Reinforcement Learning for Suspicious Subgraph Detection in Cryptocurrency Transaction Graphs
- 1Department of Computer Science and Engineering, Gyeongsang National University, Jinju 52828, Republic of Korea
- 2Department of Computer Science, Kyungil University, Gyeongsan 38428, Republic of Korea
Abstract
Detecting suspicious activity in cryptocurrency transaction networks is important for identifying illicit financial flows. When funds pass through several participants, the relationships among their transactions can reveal patterns that individual account attributes alone do not explain. Subgraph classification provides a way to evaluate these connected transactions as a group, using participant composition and local transaction paths. However, aggregating participant attributes does not preserve connection order, whereas observing only a few paths can overlook relevant information elsewhere in the group. Moreover, subgraph labels do not specify which paths are useful for detection. In this paper, we propose context-guided traversal with deep reinforcement learning to select informative paths according to the transaction group being examined. A role-wise context representation captures the composition of the group and its funding and receiving participants, enabling path selection to reflect the surrounding transaction relationships. This context is combined with the accumulated path representation to assess neighboring candidates and adapt subsequent observations. The traversal policy learns from subgraph classification performance without requiring annotated paths. Predictions from the selected paths are then combined with global transaction information to incorporate evidence beyond the visited nodes. In ten-fold cross-validation on the Elliptic2 Bitcoin dataset, the proposed method achieves average precision of 0.9657 and an F1 score of 0.9191, exceeding the corresponding mean scores of all nine evaluated comparison methods. A separate entity-disjoint test gives the highest three-seed mean AP of 0.6734 among four complete detectors. Transaction case analyses further show how the combined detector corrects missed suspicious subgraphs and false alarms from classification based on global information alone.
Related
- Dynamic Multi-Pair Trading Strategy in Cryptocurrency Markets with Deep Reinforcement Learning
- Integrating YOLOv8-based disaster detection with deep reinforcement learning for closed-loop UAV emergency response scheduling: A simulation study
- Glivenko--Cantelli Theorems for Integrated Volatility Functionals in Pure-Jump Semimartingales with an Application to Cryptocurrency Markets
- Coûts de transaction, équilibres multiples avec chômage et concurrence monopolistique
- Two-Way Price Discovery Between a Cryptocurrency-Based Prediction Market and Federal Funds Futures: Evidence from FOMC Decisions
BibTeX
@article{buu2026contextguided,
title = {{Context-Guided Traversal with Deep Reinforcement Learning for Suspicious Subgraph Detection in Cryptocurrency Transaction Graphs}},
author = {Seok-Jun Buu and Hae-Jung Kim},
journal = {Mathematics},
year = {2026},
month = oct,
doi = {10.3390/math14213924},
url = {https://www.mdpi.com/2227-7390/14/21/3924},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Context-Guided Traversal with Deep Reinforcement Learning for Suspicious Subgraph Detection in Cryptocurrency Transaction Graphs
AU - Seok-Jun Buu
AU - Hae-Jung Kim
PY - 2026
DA - 2026/10/27
JO - Mathematics
DO - 10.3390/math14213924
UR - https://www.mdpi.com/2227-7390/14/21/3924
AB - Detecting suspicious activity in cryptocurrency transaction networks is important for identifying illicit financial flows. When funds pass through several participants, the relationships among their transactions can reveal patterns that individual account attributes alone do not explain. Subgraph classification provides a way to evaluate these connected transactions as a group, using participant composition and local transaction paths. However, aggregating participant attributes does not preserve connection order, whereas observing only a few paths can overlook relevant information elsewhere in the group. Moreover, subgraph labels do not specify which paths are useful for detection. In this paper, we propose context-guided traversal with deep reinforcement learning to select informative paths according to the transaction group being examined. A role-wise context representation captures the composition of the group and its funding and receiving participants, enabling path selection to reflect the surrounding transaction relationships. This context is combined with the accumulated path representation to assess neighboring candidates and adapt subsequent observations. The traversal policy learns from subgraph classification performance without requiring annotated paths. Predictions from the selected paths are then combined with global transaction information to incorporate evidence beyond the visited nodes. In ten-fold cross-validation on the Elliptic2 Bitcoin dataset, the proposed method achieves average precision of 0.9657 and an F1 score of 0.9191, exceeding the corresponding mean scores of all nine evaluated comparison methods. A separate entity-disjoint test gives the highest three-seed mean AP of 0.6734 among four complete detectors. Transaction case analyses further show how the combined detector corrects missed suspicious subgraphs and false alarms from classification based on global information alone.
SN - 2227-7390
ER -