Continuous Embedding Spaces for Bank Transaction Data

Abstract

In the finance world, customer behavior prediction is an important concern that requires discovering hidden patterns in large amounts of registered customer transactions. The purpose of this paper is to utilize this customer transaction data for the sake of customer behavior prediction without any manual labeling of the data. To achieve this goal, elements of the banking transaction data are automatically represented in a high dimensional embedding space as continuous vectors. In this new space, the distances between the vector positions are smaller for the elements with similar financial meaning. Likewise, the distances between the unrelated elements are larger, which is very useful in automatically capturing the relationships between the financial transaction elements without any manual intervention. Although similar embedding space work has been used in the other fields such as natural language processing, our work introduces novel ideas in the application of continuous word representations technology for the financial sector. Overall, we find the initial results very encouraging and, as the future work, we plan to apply the introduced ideas for the abnormal financial customer behavior detection, fraud detection, new banking product design, and making relevant product offers to the bank customers.