Transaction Prediction by Using Graph Neural Network and Textual Industry Information

Abstract

Transaction data owned by financial institutions can be alternative source of information to comprehend real-time corporate activities. Such transaction data can be applied to predict macroeconomic indicators as well as to sophisticate credit management, customer relationship management, and etc. However, it needs attention when a financial institution uses transaction data for aforementioned applications since occurrence of transactions depends on miscellaneous factors including customer loyalty, implying missing potential transactions. To solve this issue, we can predict occurrence of transactions by formulating the problem as a link prediction task in a transaction network with bank accounts as nodes and transaction volume as edges. With the recent advances in deep learning on graphs, we can expect better link prediction. We introduce an approach to predict transaction occurrence by using graph neural network with a special attention mechanism and textual industry information, analyzing the effectiveness of the proposed method, attention mechanism and node feature design as well as demonstrating its usage as an industry importance explainer.