| Tax collection is not a new phenomenon.Every country and government will always seek out the best methods in obtaining these revenues for the use of the country.However,often times many companies evade such taxations through fraudulent activities and misleading declarations.The aim of this body of work is to bring light to a relatively understudied subject topic.This study presents a deep learning approach using autoencoder and Apriori algorithm aimed at addressing the audit challenge within the Kenya;East Africa region.Namely;the retail industry.The main purpose of this research was to develop a deep learning algorithm for tax auditing in tax fraud detection and evaluating the taxpayer’s tax compliance.This research focus was to build deep learning model for identification of risky supermarkets who fail to remit the correct taxes hence committing tax fraud.With the use of the model it will be more effective,since it will flag out risky taxpayers and this will enable to tax officers to audit them,which in return will lead to more revenue collection.This will lead to the money being allocated to the infrastructure and development of the country.Deep-learning based approach for selecting taxpayers for audit was done.The tax returns of the supermarket outlets were first transformed into a compressed representation,and then converted into the original dataset using an autoencoder.The difference between the original and the modification datasets that is the modification lapse was calculated.The modification lapse results were verified using the Apriori algorithm and the known fraudulent cases.The calculated modification lapse were consistently lower for normal observation and higher for fraudulent cases.This indicated that the approach that was taken for this research can be a good for alternative strategy for identifying tax fraud.Although the modification lapse has been frequently estimated and applied to anomaly detection in the literature,a more reliable way to verify modification lapse results was unavailable except for obtaining true cases by investing in significant resources.This research was inspired by this problem and the Apriori algorithm was used for identifying frequent patterns and same were put into consideration for providing a solution.The results showed that the algorithm can be utilized to gauge the reasonableness of the modification results created by an autoencoder.In order to run the deep learning mechanism properly python programming language was used for execution.While,the algorithm was broken down using the application of a modification lapse formula to best quantify the layers and dimensions.For this several test runs were done.The data that was collected for this paper was limited to that of major supermarket brands operating predominantly within the Kenya;East Africa region.The matter of tax declaration and collection is a very sensitive topic and such information isn’t readily available for research.However,an ample sample pool was known,retrieved,investigated and documented for this paper. |
