Classification of financial transactions in accounting using graph neural networks

Omeragić, Ermin (author), Šubelj, Lovro (mentor)

Businesses must manually assign each line item on an invoice to a particular account as part of the invoice processing procedure. When there are many account codes to pick from, this can be difficult and time-consuming. Suggesting account codes to accountants as part of the semi-automated system may reduce manual errors and speed up this process. In this thesis, we evaluated the use of graph neural networks for the task of suggesting account codes for financial transaction items. We created a heterogeneous transaction graph based on real financial data of two companies and implemented and evaluated four graph neural network models. We used methods from the previous work on this topic as baselines. The results showed a slight improvement in the performance of the graph models compared to baselines on a small-scale dataset. Further testing with more data should be carried out before the models can be deployed to a production system.

• Language: English
• Keywords: graph representation learning, machine learning, account classification, bookkeeping automation, heterogeneous networks
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-151743
• COBISS.SI-ID: 170192387
• Publication date in RUL: 18.10.2023

Secondary language

• Language: Slovenian
• Title: Uvrščanje finančnih transakcij v računovodstvu z uporabo grafovskih nevronskih mrež

Abstract

Podjetja morajo pri obdelavi računov ročno dodeliti določeni konto vsaki postavki na računu. Ko je na voljo veliko kontov, lahko to postane težavno in časovno potratno. Predlaganje kontov računovodjem kot del delno avtomatiziranega sistema lahko zmanjša ročne napake in pospeši ta postopek. V tem magistrskem delu smo ocenili uporabo grafovskih nevronskih mrež za nalogo predlaganja kontov za postavke finančnih transakcij. Ustvarili smo heterogeni transakcijski graf na podlagi resničnih finančnih podatkov dveh podjetij in implementirali ter ocenili štiri modele grafovskih nevronskih mrež. Kot osnovo smo uporabljali metode iz prejšnjih raziskav na to temo. Rezultati kažejo rahlo izboljšanje uspešnosti grafovskih modelov v primerjavi z osnovnimi modeli na majhnemu obsegu podatkov. Preden se modeli lahko uvedejo v produkcijski sistem, je treba izvesti nadaljnje testiranje z več podatki.

• Keywords: učenje omrežnih predstavitev, strojno učenje, uvrščanje kontov, avtomatizacija knjigovodstva, heterogeni grafi