Tree-based Classifiers for Smart General Ledger Code Suggestion

One of the crucial steps in accounting procedure is bookkeeping responsible for recording all the company's financial transactions. A general ledger is considered as a means of bookkeeping used for recording all the company's ongoing transactions. It is a master accounting record that provides a complete view of all the financial activities occurring in a company throughout time. Typically, the transaction data are segregated to account for assets, liabilities, owners' equity, revenues, and expenses. The objective of this study is to utilize machine learning for general ledger code suggestion, thus ensuring accurate recording of transactions to their appropriate accounts. Natural Language Processing plays an important role during the step of data preprocessing for information extraction from accounting transactions. The two techniques: TF-IDF and Bags of word, are selected as alternative vectorization for model construction. Data transformation is also performed to provide useful hints for model learning. The tree-based classifiers: CART and Random Forest, were trained to categorize eighteen classes of accounting codes. The experimental results suggest that incorporating textual information with other features could achieve higher model performance. The ensemble tree Random Forest outperformed the decision tree CART, whereas no substantial difference in the accuracy of both vectorization methods.