DeepTrax: Embedding Graphs of Financial Transactions

Abstract

Financial transactions can be considered edges in a heterogeneous graph between entities sending money and entities receiving money. For financial institutions, such a graph is likely large (with millions or billions of edges) while also sparsely connected. It becomes challenging to apply machine learning to such large and sparse graphs. Graph representation learning seeks to embed the nodes of a graph into a euclidean vector space such that graph topological properties are preserved after the transformation. In this paper, we present a novel application of representation learning to bipartite graphs of credit card transactions in order to learn embeddings of account and merchant entities. Our framework is inspired by popular approaches in graph embeddings and is trained on two internal transaction datasets. This approach yields highly effective embeddings, as quantified by link prediction AUC and F1 score. Further, the resulting entity vectors retain intuitive semantic similarity that is explored through visualizations and other qualitative analyses. Finally, we show how these embeddings can be used as features in downstream machine learning business applications such as fraud detection.

Understanding graph embedding methods and their applications

“…It is interesting to see that the lines split clearly into two distinct sets fairly quickly as training continues: a small subset of lines with low uncertainty that asymptotes a small variance constant value, and another one with increasing variance, which obviously denotes the unstable di-mensions. The authors of [4] also observed that the effective dimensionality (L = 6) observed using this approach is very close to the number of ground-truth communities (7); this is also something we observed in our own work in the two neuroscience tasks that we discuss below. 4.3.…”

“…High-dimensional and large-scale graphs can be encoded as continuous, low-dimensional graph embeddings in the latent space, while the latent space graph patterns (i.e., embeddings) can be readily used for solving diverse downstream graph analytic problems. For example, detecting an anomaly in social networks [47,12] or financial networks [7], analyzing the non-pharmacological cognitive training effects using functional brain network embedding patterns [61] and to predict early stages of Alzheimer's disease [60], mining biochemical multi-scale structures in biochemical graphs [54], finding functional modules or sub-compartments in genomic networks [1], etc. In the following, we highlight some representative applications based on four graph types: 1) social networks, 2) citation networks, 3) brain networks, 4) genomic networks.…”

Every Corporation Owns Its Structure: Corporate Credit Ratings via Graph Neural Networks

“…Cheng et al [8] proposed HGAR to learn the embedding of guarantee networks. The work developed DeepTrax [4] in order to learn embeddings of account and merchant entities. Recently, work [9] combined the spatio-temporal information for credit fraud detection.…”

“…With the advent of graph neural networks, some graph-based models [1,4,7,8,9,10] are built based on the loan guarantee network. The corporations guarantee each other and form complex loan networks to receive loans from banks during the economic expansion stage.…”

EWS-GCN: Edge Weight-Shared Graph Convolutional Network for Transactional Banking Data

“…In case of using of raw transactional data, in [9] authors propose a model based on SVM classifier, while in [3] RNN model is proposed. Recently, the idea to analyse bank clients as a part of a network was introduced in fraud detection problem [10] and in the task of embeddings construction [11], while in [12] authors solve a problem of anti-money laundering detection. Our recent work [13] also considers bank clients as a network but focuses on a different task of finding stable connections between clients which is treated as a link prediction problem.…”