Comparison of machine learning approaches for classification of invoices

Abstract

Machine learning has become one of the leading sciences governing modern world. Various disciplines specifically neural networks have recently gained a lot of attention due to its widespread applications. With the recent advances in the technology the resulting big data has augmented the need of bigger means of storage, analysis and henceforth utilization. This not only implies the efficient use of available techniques but suggests surge in the development of new algorithms and techniques. In this project, three different machine learning approaches were implemented utilizing the open source library of keras on TensorFlow as a proof of concept for the task of intelligent invoice automation. The performance of these approaches for improved business on data of invoices has been analysed using the data of two customers with two target attributes per customer as a dataset. The behaviour of neural network hyper-parameters using matplotlib and TensorBoard was empirically calculated and investigated. As part of the first approach, the standard way of implementing predictive algorithm using neural network was followed. Moreover, the hyper-parameters search space was fine-tuned, and the resulting model was studied by grid search on those hyper-parameters. This strategy of hyper-parameters was followed in the next two approaches as well. In the second approach, not only further possible improvement in prediction accuracy is achieved but also the dependency between the two target attributes by using multi-task learning was determined. As per the third implemented approach, the use of continual learning on invoices for postings was analysed. This investigation, that involves the comparison of varied machine learning approaches has broad significance in approving the currently available algorithms for handling such data and suggests means for improvement as well. It holds great prospects, including but not limited to future implementation of such approaches in the domain of finance towards improved customer experience, fraud detection and ease in the assessments of assets etc.