Auditing a business or an individual taxpayer is not today’s activity; it lasts for a long period. It was done by randomly selecting some of the taxpayers or by using a tax audit administration’s attitude. In the past, various researches have been carried out to select fraudulent taxpayers for audit. In one organization, the tax audit department is required to audit some or all of its taxpayers to check the evasion of tax and ensure compliance. There are various tax audit computational techniques that have been performed for fraud detection in recent years by tax audit administration. Auditing all taxpayers are not a very welcome procedure for the tax administration. Therefore Tax administration agencies must use their limited resources very wisely to achieve maximal taxpayer compliance, minimum intrusion, and minimum costs (Gupta, 2014) Tax compliance refer here to the taxpayers fulfilling their registration, filing, and reporting and payment obligations correctly and at the right time. In reverse noncompliance is a taxpayer that did not fulfill the above obligations. This research proposed to design and develop a hybrid machine-learning model (non-supervised plus supervised machine learning) to predict taxpayers based on their fraudulent risk level. Specifically, hybrid techniques have shown their superiorities over single techniques. A machine-learning algorithm is the best cost-effective option to make taxpayer risk-based classification and effective by developing a standardized model to classify their status. A prediction model is built using historical taxpayer data and different attributes of the taxpayer. these experiment conducted by using the KDD process model, machine learning techniques such as K-means clustering and SVM are used for clustering and classification to detect fraudulent categorical events. First, by using SVM before clustering has accuracy of 97% and, in k-means clustering, 54.63% of the records correctly clustered. Second, we got an accuracy of 99% by using SVM after clustering.