Neural Networks and Deep Learning

Lecture 9 · Game: Judge the transaction by its neighbours

Judge the transaction by its neighbours

You work in the anti-money-laundering team of a crypto exchange. Each card shows a real Bitcoin transaction from the Elliptic graph (the test period of Lecture 9, time steps 35–49) and the transactions linked to it by payment flows, up to two steps away. Neighbours already known to be illicit are red, licit ones green, unknown ones grey. Decide: illicit or licit? Among all test transactions 6.5% are illicit; in the game 40% are, so that you meet enough of them.

You play against the graph sample and aggregate network (GraphSAGE) of the lecture, a graph neural network (GNN) trained on time steps 1–34. It never sees the labels of the neighbours, only their 165 features, which it averages over two steps of the graph. It flags a transaction as illicit when its probability is at least 50%.

Class board for the projector

Data: Elliptic Bitcoin transaction graph (Weber et al., 2019; 203,769 transactions). The 30 transactions of the game are drawn at random from the labelled test transactions with 1–12 direct neighbours, at least one of them labelled (12 illicit, 18 licit; seed 20261108); every game shows 4 illicit and 6 licit. Per neighbour at most 6 second-step neighbours are drawn, labelled ones first; the others are counted. Predictions: mean probability of the three GraphSAGE and MLP networks of the lecture (lectures/L9/cache/models.npz). Code: lectures/L9/make_game_L9.py.