Predictive Analytics as a Service on Tax Evasion using Gaussian Regression Process

Author Name(s): S.Kishore Babu, *S.Vasavi
Author Email: skbjpj@gmail.com

Abstract

Predictive analytics combines the capabilities of statistical analysis, machine learning and data mining. Vast amount of unstructured data produced by various public and private sectors such as government, health insurance, social media and academics gave the way for text analytics to make an insight into finding risk. Predictive analytics can forecast trends, determines statistical probabilities and to act upon fraud and security threats for big data applications such as business trading, fraud detection, crime investigation, banking, insurance, enterprise security, government, healthcare, e-commerce, and telecommunications. Predictive analytics as a service (PAaaS) framework is proposed in our earlier works. This paper gives solution to one of the application fraud detection in income tax data. The solution is based upon ensemble model that uses Gaussian process with varying hyper parameters. Performance measures NRMSE and COD are used to analyse the model. Test results proved that the third hyper-parameter values yielded a good result with less error rate and more variance which is reliable for a predictive model.

Keywords

Predictive Analytics, Gaussian process, Data Transformation , Statistical Computing, Tax Administration

Introduction

Data analytics uncover hidden patterns and correlations from large volumes of data by using techniques from statistics, machine learning, artificial intelligence and data mining. Predictive analytics (PA) refers to predictions such as customer relationship management, cross sell, healthcare insurance, risk management in banking, telecommunications, fraud detection about the future through analysis of data.  Data analytics can be categorized into descriptive, predictive and prescriptive models. Descriptive models uses data aggregation to conclude upon what has happened, like mentioning the relationship between the data and describes past, predictive analytics that uses statistical models to forecast the future like what may happen and prescriptive analytics uses optimization techniques to suggest the ways of outcomes and their possible effects. Further predictive models can be categorized into classification models and regression models. Classification models determine class labels (categorical) where as regression models help in predicting a numeric value. Many techniques have been developed for predictive modeling such as SVM, Bayesian methods, neural networks, regression models, k-NN, uplift models and decision trees. But ensemble models proved to be achieving good accuracy when compared to others, reason being, they train several similar models and combines results so that a best model can be derived to predict new data. As explained in [1], predictive models can find relationship between outcome and dependent variables.Similarly descriptive models are used to form clusters of objects with similar characteristics. There are six phases for predictive analytic process. In the initial phase project is defined with outcomes, objectives, scope and the deliverables from the project. In the next phase data is collected from various sources and is analyzed. This analysis requires strategies for preprocessing such as data cleaning, transformation and data modeling so that useful data is extracted for further processing. Subsequently validate the initial hypothesis using statistical models. The next phase is predictive modeling for forecasting the future. Results after implementation can be deployed for using it in the day to day decision making.  The last phase is monitoring the model in order to ensure that it is providing the expected results.

Conclusion

This paper presented one of the case studies, where we build a predictive model that forecast the tax values. Regression has been used successfully in building the predictive model by selecting suitable hyper-parameters that are used in defining Gaussian functions. The performance of the predictive model has been measured by using Normalized Root Mean Square Error (NRMSE) and Coefficient of Determination (COD) performance measures of different hyper-parameter values such as ,  and . It was observed that when the value isperformance is good when compared to.

The field of Artificial Neural Networks, Gaussian process, Linear Regression model has diverse opportunities for future research in the predictive analytics. This paper presented results of Gaussian Process prediction technique to build the predictive model. In future, work can be extended by implementing various other techniques such as Artificial Neural Networks and Linear Regression Model techniques. By implementing all these techniques, we can determine which technique is more reliable by using performance measures such as NRMSE and COD.

439 total views, 4 views today

Download PDF File

About the author: tej