Detecting Anomalies In Value For Duty Purposes(VDP) In Imported Motor Vehicles

Tax evasion and tax fraud have been a constant concern for tax administrations in Revenue Authorities.In order to combat these problems and improve revenue collection,there has been a growing emphasis on proactive Risk Management System(RMS).These systems are designed to identify and mitigate risks before calculation of taxes.Tax revenue,particularly in Malawi is very important and contributes to more than 50%of the government revenue.For this reason,it is important to know factors that influence the collection of proper and accurate taxes.This research,therefore proposes an approach for identifying anomalies in customs transaction data through a careful analysis of the Value for Duty Purposes(VDP).Anomaly detection has been used extensively in a wide range of real world problems and has attracted significant attention in a number of research areas over the last decade.It has attempted to recognize events,activities,or observations that are measured differently than the expected behavior or pattern obtainable in a dataset.The most widely used anomaly detection techniques are cluster analysis and classification.The study uses cluster analysis and classification for anomaly detection.The study focuses on detecting anomalies in Value for Duty Purposes(VDP)on imported motor vehicles by using the supervised learning approach in which anomaly detection process begins by applying the approach to differentiate normal behavior classes(context)before attempting modeling normal behaviors.Different classification and clustering methodologies have been used to classify and cluster the anomalies.The WEKA data mining tool was used by employing the J48,Logistic,Decision Table,Interquartile Range,ZeroR and K-means algorithms for evaluation.The Statistical Package for Social Scientists(SPSS)was also used in the analysis to generate boxplots to show out-of-range instances.The study used data from the Malawi Revenue Authority(MRA).A total of 7097 transactions were picked from the MRA database.Before building the clustering and classification models,these transactions were preprocessed by applying normalization procedures.After preprocessing distinct experiments were conducted by performing adjustments on the attributes in order to come up with purposeful outputs.The results from SPSS show that most of the variables are significantly correlated.The variables which are positively correlated with the target variable,VDP,included year of make,vehicle description,FOB and custom duty.Further analysis in SPSS using Boxplots was able to pick outliers and a check on the outliers confirmed whether an item is anomalous or normal.Classification results from WEKA using ZeroR for baseline performance assessment gave an accuracy rate of about 80%.Further classification analysis using J48 gave an accuracy rate of 90.4%and Logistic Regression had an 88.7%accuracy rate.When the decision tables were used to classify the instances,the results indicated an 89.9%accuracy rate.Clustering results using K-means split the data into five clusters.One class labeled C3 had 1585 anomalous instances which represented 22%of the sample.Additional analysis using the interquartile range(IQR)was able to pick extreme values and outliers.A focus on the interquartile range showed that 12.4%of the 7097 instances(i.e.879)were outliers and 42.5%of the 7097(i.e.3017)instances were extreme values.Overall,this research has verified that data mining techniques can be used in detecting anomalies in VDP and boost tax collections.All the experiments gave valid results and can be used to detect anomalies.However,a comparisons of the experiments was done to determine which experiments performs better than others.J48 classification and 1QR performed better than the rest.Therefore this study recommends the adoption of J48 and IQR.In summary,this study makes several contributions both academically and practically.First,the thesis proposes a machine learning framework to detect anomalies in VDP on imported motor vehicles,including data collection and processing,attribute relevance analysis,feature design and formalization,and anomaly detection algorithms.Secondly,a number of experiments were conducted,including classification,clustering,and interquartile range approach.A comparative analysis with various kinds of machine learning algorithms achieved good results and conclusions.Lastly,a prospect of the application of the proposed framework has been presented which is believed to improve the efficiency of the customs management system in Malawi.