Granularity Effects on AML Queue Composition in Elliptic++ Bitcoin Graph
A new paper on arXiv (2604.23494) evaluates how transaction-level versus actor-level anti-money laundering (AML) scoring affects investigation queue composition on blockchain networks. The authors propose a projection framework with four aggregation operators to map transaction scores to actor-level action units, introducing budgeted metrics: yield@budget, burden decomposition, and case fragmentation. Using the Elliptic++ Bitcoin dataset (203,769 transactions; 822,942 address occurrences), they train independent random forest classifiers at each granularity level under a causal temporal protocol. At a one-percent budget, they compare queues via Jaccard overlap, burden decomposition, and feature-matching ablations. The study highlights that granularity choice significantly impacts queue composition, with implications for compliance efficiency.
Key facts
- Paper arXiv:2604.23494 evaluates transaction-level vs actor-level AML scoring granularity
- Elliptic++ Bitcoin dataset contains 203,769 transactions and 822,942 address occurrences
- Four aggregation operators map transaction scores to actor-level action units
- Budgeted metrics include yield@budget, burden decomposition, and case fragmentation
- Random forest classifiers trained under causal temporal protocol
- Comparison at one-percent budget using Jaccard overlap and feature-matching ablations
- Granularity choice significantly affects investigation queue composition
- Study has implications for compliance efficiency in AML systems
Entities
Institutions
- arXiv