US 12,321,943 B2
Locating suspect transaction patterns in financial networks
Andrea Giovannini, Zurich (CH); and Tamas Visegrady, Zurich (CH)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Jan. 4, 2022, as Appl. No. 17/568,001.
Prior Publication US 2023/0214842 A1, Jul. 6, 2023
Int. Cl. G06Q 20/40 (2012.01); G06N 3/08 (2023.01)
CPC G06Q 20/4016 (2013.01) [G06N 3/08 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
generating a first transaction graph for a first financial network;
training a generative graph neural network to generate synthetic suspect pattern graphs based on a seed set of suspect transaction pattern graphs; and
generating a plurality of synthetic suspect transaction pattern graphs based on the trained generative graph neural network;
comparing a pattern of transfer amounts for nodes in the synthetic suspect pattern subgraph to a first set of patterns of transaction values for nodes in the first transaction graph;
identifying one or more corresponding transaction patterns in the first transaction graph; and
modifying the identified one or more corresponding transaction patterns in the first transaction graph to match the synthetic suspect pattern subgraph;
extracting a first set of one or more subgraphs from the first transaction graph, wherein each of the one or more subgraphs of the first set is comprised of a randomly-selected node and a group of nodes reachable from the randomly-selected node via corresponding edges in the transaction graph;
training a graph neural network model to classify a subgraph as suspect with the first set of the one or more extracted subgraphs;
generating a second transaction graph for a second financial network using a set of new transaction data;
extracting a second set of subgraphs from the second transaction graph;
inputting the second set of subgraphs into the trained graph neural network model for classification of each subgraph;
classifying, by the trained graph neural network model, a first subgraph of the second financial network as suspect; and
responsive to the first subgraph being classified as suspect, freezing accounts corresponding to nodes in the first subgraph.