Abstract:
The collection of tax is the mainsource of incomefor the government.
Taxcollecting has been associated with a lot of fraud, which is a challenge to
detect.Fraud involves one or more persons who intentionally act secretly to
deprive thegovernment of income and use it for their benefit. This study was initiated
to
explorethedeeplearningtechnologyfordevelopingmodelsthatcandetecttaxfraudusing
dataobtained fromtheMinistryofRevenuesinEthiopia.
To collect the data, the researcher used interviews and observation as primary
dataand database analysis as secondary data. The dataset used in this study had
beentakenfromEthiopia's Ministryof Revenues.Afterselectingthedataset,pre processing techniques such as filling missing records, removing outliers,
reducingthe dimension, selecting the most relevant features, and finally
normalizing thedataset input using features scaling are performed. The deep learning
models for
taxfrauddetectionareimplementedusingPythonprogramminglanguage.Theexperime
nts had beenconducted by using the 23536-dataset records.We used 80%of the dataset
for training the model and the remaining 20% of the dataset for testingthe
performance of the model that is developedby the ConvolutionalNeuralnetwork.
The model had shown the highest classification accuracy of 84.64%. Thenthis model
was tested by 4708 testing datasets and scored a prediction accuracy of84.41%.
The results of this study have shown that deep learning technology
isvaluablefortaxfrauddetection.