Network-based Active Learning for Identifying Illicit Actors in Financial Transaction Networks

Amro Alabsi Aljundi, Abhijin Adiga, Christopher Barrett, Margaret J. Foster, Brian D. Klahn, Dustin Machi, Achla Marathe, Philip B.K. Potter, Aaron Schroeder, Samarth Swarup, Anil Vullikanti, Mandy L Wilson, Madhav Marathe

Keywords: Active learning, collective classification, cryptocurrency, financial networks

Abstract:

Identifying illicit transactions within financial networks, such as
Bitcoin, is an important area of research. Available datasets are
highly imbalanced, making the design of machine learning methods
challenging. Active learning, which carefully chooses data points
for annotation, has been shown to improve performance for such
problems.
Here, we design a new approach, C2AL, for detecting illicit nodes
in financial networks, which incorporates network correlations
more explicitly. Our approach builds on prior work on active learn-
ing on networks, specifically, collective classification (CC), which
uses predicted labels of neighboring nodes to improve classification.
We extend this approach by incorporating the information from
both underlying models of collective classification, as well as their
contrastive information, into the active learning sample selection
procedure. We show that C2AL significantly improves sample effi-
ciency, requiring 24–48% fewer labeled samples than prior methods
to achieve comparable detection performance across six financial
network datasets.

Area: Learning and Adaptation (LEARN)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1433