Enhancing bookkeeper decision support through graph representation learning for bank reconciliation

The emergence of cloud-based bookkeeping platforms has made it possible to streamline decision-making in tedious accounting tasks, such as bank reconciliation. Bank reconciliation involves tracing the expected cash flow from invoices and bills to the actual payments recorded in a business’s bank feeds. A bookkeeper is responsible for ensuring that every payment listed in a business’s bank statement is accurately matched to financial activity recorded in the bookkeeping system. This process is crucial for maintaining the accuracy and integrity of the business’s financial records. Current decision-support systems leverage natural language processing to recommend close matches for incoming bank feeds. While these approaches are effective for one-to-one matching, they underperform in identifying one-to-many matches, which are common and significantly more complex for larger businesses. In this work, we investigate the value of embedding relational data along with natural language in identifying matches to support the bank reconciliation process. Our proposed graph-based system surpasses industry benchmarks on one-to-one matching and offers a more robust decision support solution for the identification of one-to-many matches. Additionally, we introduce a novel post-processing technique, Top Boundary Ranking, which enhances the system’s detection of group-based matches.