Graph Analytics—Powering the Game Against Money Laundering

Govind Gopinathan Nair, Technical Product Manager, Financial Crime and Compliance | January 24, 2022

The rapid adoption of digital payments and proliferation of financial technology services has expanded access to financial services globally and made it easier and faster to move money across the world. This has led to an explosive increase in the volume of financial activities and accompanying data. These large troves of data contain signals that can enable the detection and prevention of financial crime.

Most anti-money laundering (AML) solutions in use today are rules-based behavior detection systems that are not designed to identify complex, suspicious patterns of transactional activity. Moreover, traditional AML systems often use relational databases; determining relationships and connectedness between entities in such relational databases can be challenging. This is a major limitation that can impede efforts to detect and investigate potential money laundering cases.

An evolving technology that can give financial institutions a leg up in the fight against financial crime is the combination of graphs and graph analytics. Graphs and graph machine learning are already being used to build knowledge graphs that power recommendations and search in some of the largest technology companies, to power cutting-edge research to solve protein folding and to discover new drugs. Similarly, graph and graph analytics can be a potent tool in the financial crime-fighting tool kit in the financial industry.

Understanding graph analytics

In its simplest form, a graph consists of nodes or vertices that represent entities connected by edges representing the relationships between these entities. These graphs can be directed or undirected depending on the nature of these relationships. Graphs are the most natural way of capturing relationships. Furthermore, property graphs allow additional data about nodes and edges to be captured as the node and edge properties.

Modern graph analytics tools can analyze the relationships between entities, attributes of these entities as captured by node and edge properties, and how these evolve over time. Graph query languages like PGQL can allow users to precisely query for complex patterns. Additionally, modern graph neural networks can learn representations of such graphs that combine the topology and relationships of graphs along with their node and edge properties. As such, graph analytics is emerging as the tool of choice to analyze relationships, complex dependencies, hidden linkages, networks, and clusters.

Financial entities, transactions, and relationships can be naturally represented using graphs. Graph analytics allows AML applications to identify, query, analyze, and visualize relationships; based on transactions, shared addresses, phone numbers, or e-mails; between entities such as customers, accounts, households, etc. Thus, it is quickly becoming a sought-after tool by analysts and AML practitioners to mine financial transaction data for insights and understand complex, non-obvious relationships.

Leveraging graph analytics to combat money laundering

The ability to analyze links and relationships between entities makes graph analytics the ideal tool for AML. Graph analytics presents several opportunities to combat money laundering innovatively. It can improve the efficacy and effectiveness of a traditional AML program multi-fold. Let us see how.

1. Identify higher risk entities

Ranking algorithms such as closeness centrality, degree centrality, eigenvector centrality, etc., can be used to rank nodes in a graph. These measures capture the importance of a node to a graph along different dimensions.

For instance, degree centrality captures how connected each node is in a graph, whereas eigenvector centrality measures how connected a node is to other highly connected nodes in the graph. Such centrality measures can determine the most significant nodes in the financial graph.

Degree distribution algorithms are an easy way to analyze the structure of a graph. For example, in a typical transaction graph, entities with the highest vertex degree (number of neighbors) are usually business entities. Institutions can analyze the degree distributions of their customers and identify outliers with unusually high degrees given their customer profile. Such entities might be candidates for enhanced due diligence or ongoing due diligence as part of ongoing know your customer (KYC) process.

2. Enhanced detection and monitoring

Graph querying languages like PGQL allow users to write queries or scenarios that capture complex patterns of fund movements. Such tools allow for more tailored monitoring for specific high-risk patterns. This can be particularly useful for identifying ultimate beneficiary owners (UBO), where these UBOs are embedded in a complex chain of ownership and transactions.

Graph algorithms can be used to find the shortest path between nodes in the non-transaction graph (graph considering only non-transactional relationships). If the shortest path in the transaction graph (considering only transaction data) between the same nodes is much longer, it might indicate an attempt to layer funds.

Modern graph neural networks also allow us to learn embeddings or representations of the nodes in a graph. The embeddings capture the topology, relationships, and properties of a node. Such embeddings can also be used in downstream models such as customer risk scoring or event scoring that can greatly improve models’ performance, reducing both false positives and false negatives. Graph Neural Network explainers are also available that can address concerns around explainability of these embeddings.

3. Provide context to and accelerate investigations:

Whenever an alert is flagged, it is important for the AML analysts to determine whether this is an isolated or interconnected incident. In a traditional AML investigation, it would be difficult to identify connectedness in the scattered datasets (customers, accounts, transactions, etc.). However, constructing a graph to represent a case enables graph visualizations and analytics, helping investigators get a contextual view of the investigated entity.

Modern graph deep learning techniques also allow us to learn embeddings for the cases and then surface similar suspicious activity reports (SAR) that can provide useful guidance to investigators.

4. Financial crime knowledge graphs

A long-term goal for financial institutions can be to construct a financial crime knowledge graph. Combining modern natural language processing (NLP) and graph databases will allow institutions to create a single financial crime graph that captures all structured and non-structured, internal and external data on customers. This will enable a deeper understanding of customers which will be useful across various functions such as KYC, investigations, and even marketing.

Bringing it all together

This is just a sampling of the potential use cases graph analytics enables. Institutions can experiment with some simpler use cases before embarking on large-scale adoption.

Graph analytics can empower data scientists to identify anomalies and patterns that can improve detection, reduce costs and deliver faster time to AML compliance. It also offers powerful visualization capabilities that can markedly improve investigator productivity and help them to understand complex intricate activity patterns.

Implementing graph analytics as part of the AML toolkit would need skilled resources, investment, and commitment; however, the benefits outweigh these costs as embracing graph analytics can turbocharge AML compliance programs at banks and financial institutions.