Artificial Intelligence Systems for Value Added Tax Collection via Self Organizing Map (SOM)

doi:10.21203/rs.3.rs-2216885/v1

Download PDF

Research Article

Artificial Intelligence Systems for Value Added Tax Collection via Self Organizing Map (SOM)

https://doi.org/10.21203/rs.3.rs-2216885/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Purpose: Corporates and private businesses, primarily uses Artificial Intelligence to influence business models, sales processes, customer segmentation, strategy formation and as well as to understand customer behaviours, thus increase revenue. There is substantial research on the influence of Artificial Intelligence on business strategies with the objectives of increasing revenue (Stallkamp & Schotter, 2021). However, there is limited research on the use of AI in information systems research to assist in the efforts of revenue collection and VAT fraud detection. Subsequently, the present study explores the use of Artificial Intelligence in VAT fraud detection. The main purpose of the study is to determine how corporate VAT fraud could be detected in real time (Alshantti and Rasheed, 2021). The research question poses to identify what type of AI technique or framework could be applied to determine VAT fraud?

Problem Statement: VAT fraud is a major problem for tax administrations across the world. Tax fraud detection is likely to become even more important with recent developments in Artificial Intelligence. Auditing VAT declarations is a slow and costly process that is very prone to errors. Conducting VAT audits for example, involves costs to the tax administration as well as to the taxpayer. Furthermore, the field of fraud detection is characterised by unlabelled historical datasets. In this study the use of unsupervised machine learning algorithm is put forward. Unsupervised algorithms are well suited to unlabelled historical datasets, common in the fraud detection or classification arena. The authors conduct experiments using an unsupervised Neural Network algorithm to classify suspicious Value Added Tax declarations. Thus, assist in the efforts of tax audits by tax administrations. Consequently, it is envisaged that the chances of detecting fraudulent VAT declarations will be enhanced using AI techniques as proposed in this research.

Methods: Neural Network Self-Organizing Map (SOM) methodology based on VAT fraud detection.

Results: Using SOM methodology, the results shows that VAT fraud or suspicious behaviour can be detected or differentiated by observing VAT declarations from attributes such as VAT liability, exempt supplies, refund, and input VAT on capital goods purchased.

Conclusion: In conclusion, the value added tax credit (VAT) and refund mechanism offers unique opportunities for abuse, and thus has recently become a crucial concern for revenue services (e.g., Internal Revenue Service, the South African Revenue Service (SARS) and HM Revenue and Customs). Consequently, this study presented artificial intelligenceframework that detects VAT fraud in real-time.

Self Organizing Map

Anomaly Detection

Artificial Intelligence

ICT impact

Robotic Process Automation

Analytics and Data Science

In information Systems field, IS or IT business strategies and modelling could be referring to as an act or science of initiating a transaction or exchange through a predetermined series of actions (e.g., organizational management, planning or technology processes). Research on digital platform (or multisided markets) originated in IS economics and, has been pronounced in strategy field since early 2000s (McIntyre & Srinivasan, 2017; Stallkamp & Schotter, 2018). The adoption of Internet and mobile phone services has enabled industries to introduce a platform enabled business models commonly referred to as disrupting industry structure (Hagia, 2009; Bankole and Bankole, 2017). For instance, in transportation, lodging and meal delivery sectors (such as Uber, Mr. delivery and Airbnb).

Digital platform (e-platform) is an ICT value creation which facilitates transactions between several groups of users including buyers and sellers (Stallkamp & Schotter, 2018). For example, content and search engine optimization, social media marketing and optimization are augmenting consumer buying powers as more and more consumers are voicing their opinions. Furthermore, customers express their views about the industry, their brands and related product attributes. Ineffective delivery of product and services regarding customer requirements could impact the corporate brand, image, loyalty, and values (Mohlala and Bankole, 2021). This could cause customer discontent thereby disengagement with products and brands through eWom (Huete-Alcocer, 2017). Overall, digital technologies present considerable opportunities for enterprise leaders to rethink their business to create better experience for customers, employees, partners, and as well lower cost of services (Bendor-Samuel, 2018). In this research, we explore how an AI digital platform framework could be employed to explore value added tax fraud prediction in revenue service sector.

Notably, artificial intelligence (AI) digital platform appears to influence effective and productive revenue collection strategic decisions (Abdel-Karim, Pfeuffer and Hinz, 2021). The recent work and trends in field of AI digital platforms varies namely, Apple Siri enabled smart mobile searches, the web search and capture of keywords, google duplex for hair-grooming appointment, restaurant reservations, voice tone and language patterns duplex are hardly distinguishable with human voice (Haselton, 2018). Recently, Amazon entered a partnership with Marriott International Inc. wherein Amazon Flywheel and Amazon Alexa Voice enabled platforms performs the task of assisting hotel guests from room servicing to housekeeping (Dignan, 2018). Regardless, we alluded that few research existed around the use of AI in computer information systems that explores digital platforms in aiding the efforts of revenue collection and leakages thereof, due to tax fraud and evasion.

VAT fraud can be explained as a deliberate misrepresentation of information in VAT returns or declarations to decrease the amount of the tax liability (Vanhoeyveld, Martens, & Peeters, 2019). VAT fraud is a major problem for tax administrations across the world. It is carried out by criminals and organized crime networks. VAT fraud can occur in many sectors including electronics, minerals, cars, and carbon permits. The most attractive goods for fraudsters have been those of high value and low volume such as mobile phones or computer chips, which generate huge amounts of VAT in the lowest number of transactions and in the shortest possible time (Madongo, 2017). At the heart of the VAT system is the credit mechanism, with tax charged by a seller available as a credit against their liability on their own sales and, if more than the output VAT due, refunded to them. According to (Keen & Smith, 2007), this creates opportunities for several types of fraud characteristic of the VAT namely: False Claims for Credit or Refund; Zero-rating of Exports and Misclassification of Commodities; Credit Claimed for VAT on Purchases that are not Creditable; Bogus Traders; Under-reported Sales; Failure to Register; Tax Collected but not Remitted.

The development of AI e-platform for VAT fraud detection is required to ensure that large amounts of revenue that could be used by the government for the much-needed socio-economic public services such as hospitals, schools and road infrastructure are generated. Artificial neural network (ANN), when trained properly can work like a human brain. They learn by example, like people and are known as an exceptionally good classifier. Furthermore, the neural network is preferred in this study due to its ability solve classification problems (Georgieva, Markova, & Pavlov, 2019). Machine Learning algorithms are very likely to produce faulty classifiers when they are trained with imbalanced datasets. Fraud datasets are characterized by imbalanced datasets. An imbalanced dataset is one where the number of observations belonging to one class is significantly higher than those belonging to the other classes. Other algorithms tend to show a bias for the majority class, treating the minority class as a noise in the dataset. In many standard classifier algorithms, such as Naive Bayes, Logistic Regression, and Decision Trees, there is a likelihood of the wrong classification of the minority class. ANN are well suited to imbalanced datasets (Kelechi, 2018). Hence, this research proposes a Self-Organizing Map (SOM) Neural Network algorithm to detect VAT fraud.

A self-organizing map (SOM) or self-organizing feature map (SOFM) is a type of artificial neural network (ANN). The SOM is trained using unsupervised learning to produce a low dimensional map. It is a discretized representation of the input space of the training samples, called a map. Self-organizing maps differ from other AI algorithms in that it uses competitive learning instead of error-correction learning (Kohonen, 1982). In a sense, it uses a neighbourhood function to preserve the topological properties of the input space (Bansal & Suman, 2014).

The background for this research is multifold, that is, to create a VAT Fraud detection AI Framework and the application of a Self-Organizing Map (SOM) algorithm to detect VAT fraud. This paper focuses on VAT fraud detection. As mentioned in Section 1 above, the fraud detection arena is characterised by extremely large amount of unlabelled structured and unstructured data. Unsupervised machine learning algorithms are well suited to unlabeled datasets. Hence, herein we propose an unsupervised machine learning approach for detecting VAT fraud. We begin by describing the SOM algorithm. Thereafter, we discuss data collection and data preparation (Vanhoeyveld, Martens, & Peeters, 2019). Additionally, we elaborate on issues relating to VAT variables feature selection, followed by an exploratory data analysis. Furthermore, we explain the machine learning technique and statistical algorithm employed in this study. Finally, we present results from actual digital platform-based experiments conducted on taxpayer level VAT dataset.

2.1 Context

In the meantime, researchers have made immense strides in recent years in machine learning augmentation. Machine learning augmentation is concerned with the systematic study of algorithms and systems that improve their knowledge or performance with experience (Flach, 2012). Machine learning applications are ubiquitous within data science. For instance, regression, clustering, and classification are three important techniques within the data science space (Flach, 2012). However, their application is not limited to aforementioned areas. Although, machine learning application is traditionally used in the data-modeling step of the data science process (Flach, 2012). Nonetheless, technology integration into business domains is not a new phenomenon which means, a digital platform function incorporates significant technology integration into conventional operations and presents an added advantage for brand promotion. In the surmise of Molsa (2017), Artificial Intelligence & automation are poised to reshape the digital platform function.

Phua, Alahakoon and Lee (2004) in their paper, tabulate, compare and summarize fraud detection methods and techniques that have been published in academic and industrial research during the past 10 years. This is done in the business context of harvesting the data to achieve higher cost savings. Phua et al (2004) presents a methodology and techniques used for fraud detection together with their inherent problems. In their research, they juxtaposed four major methods commonly used for applying a machine learning algorithm. Firstly, supervised learning on labelled data. Secondly, hybrid approaches with labelled data. Thirdly, semi-supervised approach with non-fraud data, and lastly, unsupervised approaches with un-labelled data. (Phua, Alahakoon & 2004). Meanwhile, Shao et al. (2002) describe the building of a fraud detection model for Qingdao customs port of China. The model is used to provide decision rules to the Chinese custom officials for inspection of goods based on historical transaction data. The objective is to improve the hit rate. The model is appropriately named ‘Intelligent Eyes’ and has been successfully implemented with high predictive accuracy (Shao, Hong, & Chang, 2002).

Nagadevara & Kumar, (2006) have developed a risk assessment predictive algorithm for selective customs examinations in the Indian customs authority. The model was developed using a classification tree algorithm and was expected to detect over 90% of the total duty short declarations. The model reduced the original manual examination efforts to the third of the time (Nagadevara & Kumar, 2006).

Notwithstanding, Tax administration agencies must therefore use their limited resources very judiciously whilst achieving maximal taxpayer compliance albeit at the lowest cost of revenue collection. Whilst, at the same time, adhering to lower levels of taxpayer intrusion.

The Quantitative Analytics Unit of the Securities Regulation Institute in Coronado, California, USA, developed a revolutionary new statistical-based algorithm application called “NEAT,” which stands for the “National Examination Analytics Tool” (White, 2014). With NEAT, securities examiners can access and systematically analyse massive amounts of trading data from firms in a fraction of the time it has taken years in the past. In one recent examination, NEAT was used to scrutinize in 36 hours exactly 17 million transactions executed by one investment adviser (White, 2014). Among its many benefits, NEAT can search for evidence of probable insider trading by comparing historical data of significant corporate activity like mergers and acquisitions against the companies in which a registrant is trading. They then use this information to analyse how the registrant traded at the time of those notable events. NEAT can review all the securities the registrant traded and quickly identify the trading patterns of the registrant for suspicious activities (White, 2014).

2.2 Theoretical background

2.2.1 Artificial Intelligence (AI) Digital Platform

AI e-platform uses artificial intelligence techniques to make automated decisions based on data collection, data analysis and data scrutiny. AI digital platform serves as a Computer Information Systems platform that showcase economic trends that may impact system automation efforts. Artificial Intelligence techniques like machine learning use customer data, to learn how to best interact with customers thereby providing insights that could serve those customers with tailored messages at the right time without intervention from external factors so as to guarantee effective, efficient and impactful product development and communication (Larsson, 2021).

In the current circumstances, this study endeavored to place AI digital platform development in the context of systemic developments which could be well-thought-out as digitalization of industries IT resources (Larsson, 2021).

Additionally, AI e-platform perform repetitive, routine, and tactical tasks that require less human intervention. It use-cases may include data analysis; media buying; automated decision making; natural language processing; generation of content; real-time personalized or tailored messaging (Larsson, 2021). Accordingly, artificial intelligence digital platforms hold a vital role in helping managers to understand machine learning algorithms like k-nearest neighbor, Bayesian Learning and Forgetting, Self-Organizing Maps, Artificial Neural Network Self-Organizing Maps. These forms of algorithms help to gain a comprehensible understanding of how amenable and responsive a customer is to a specific product offering effort.

Therefore, AI e-platform frameworks are required to process expansive and extensive data sets that would unveil hidden knowledge and insights about product and their customers. Thus, enabling organizations to derive significant revenue growth, whilst strengthening customer relationships (Marketing Evolution, 2021). There is significant research on the impact of artificial intelligence on business process in information systems to increase revenue, however more research is needed to explore its potential in aiding revenue collection by tax

administrations (Smith, 1970). Consequently, in this current study we use unsupervised models to detect Value Added Tax fraud in order to improve tax compliance, and thus enhance revenue collection.

2.2.2 Social Behaviours on Tax Fraud and Compliance

Earlier on, we explained that AI e-platform is required to understand customer needs to create appropriate and personalized messages and product offering. In the same vein, tax administration’s understanding of its taxpayers is a key to effective tax administration and revenue collection. The taxpayers’ attitude on compliance may be influenced by many factors, which eventually influence taxpayer’s behavior. Those factors which influence tax compliance behavior differ from one country to another, and from one individual to another (Kirchler, 2007). Namely, taxpayer’s perceptions of the tax system and Tax Authority (Ambrecht, 1998); peer attitude, norms and values; taxpayers ’s understanding of the tax system or tax laws (Silvani, 1992); motivation such as rewards (Feld, Frey, & Targler, 2006); punishment such as penalties (Allingham & Sandmo, 1972); cost of compliance (Slemrod, 1992); enforcement efforts such as audit; probability of detection; difference across culture; perceived behavioral control (Furnharn, 1983) ; ethics or morality of the taxpayer and tax collector; equity of the tax systems; demographic factors such as sex, age, education and size of income and use of informants (Murphy, 2004).

Notwithstanding, tax fraud detection, enforcement and the behavior of others affect taxpayer compliance (Weisman, 2001). The IRS Commissioner Charles Rossotti noted that when the number of audits is reduced, honesty suffers as fears of policing decline. Additionally, if taxpayers begin to believe that others are cheating, then the temptations to shave their own tax burdens may become irresistible. Commissioner Rossotti's observations recognizes that tax fraud detection, enforcement affects social behaviors, and that these behaviors can, in turn, affect taxpayers' compliance decisions. (Weisman, 2001). Accordingly, the probability that a taxpayer will escape their tax obligations increases when the taxpayer suspects that his associates, colleagues, and acquaintances are evading taxes (Bird & Milka, 1992).

3.1 Overview of the Framework

The VAT Fraud Detection AI framework serves as e-platforms that employs a Self-Organising Map (SOM) neural network. The framework is designed with the objective of assisting with the VAT audit case selection. The proposed framework would improve the effectiveness and efficiency of the revenue collection agencies in identifying anomalies on VAT returns filed by the taxpayers. The framework hereunder, classify and segregate taxpayers into clusters or categories that have the greatest likelihood of committing fraud. Thus, the framework selects taxpayers for audit on the probability that they have committed fraud. The VAT Fraud Detection AI Framework proposed herein, is an amalgamation of a typical industry standard machine-learning life cycle and tax authorities’ VAT auditors’ standard guide.

3.2 Flow of the Framework

Task 1 – Extract VAT return data

According to the industry standard machine learning life cycle, this task is conducted under the data gathering phase of the life cycle. This step is involved with the collection of data, identification of other data sources and to integrate the data obtained from various sources. The goal of this step is to identify and obtain all data-related problems. In this step, we need to identify the different data sources, as data can be collected from various sources such as files, database, internet, or mobile devices. It is one of the most important steps of the life cycle. The quantity and quality of the collected data will determine the efficiency of the output. The more data we collect, the more accurate will be the classification or prediction.

Task 2 – Aggregate Data

After collecting the data, we need to prepare it for further steps. In the machine learning life cycle this task is completed under the Data preparation phase. Data preparation is a step where we put our data into a suitable database or files and prepare it to use in our machine learning training. In our framework, for each VAT dealer, we aggregate all numerical continuous variable obtained from the return. In this study the summary values, a period of 6 years, and are calculated for each individual VAT vendor. This effectively allows the algorithm to have a work with view of the vendor behaviour as opposed to monthly or yearly scrutiny. During this task will also conduct data pre-processing and exploratory data analysis.

Task 3 – Normalize Data

This task is normally undertaken under the Data preparation phase of the Machine Learning Lifecycle. During data preparation we use a technique called normalization or standardization, to rescale our input and output variables prior to training a neural network model. The purpose, of scaling the dataset, when training a Neural Network model, is to normalize the data to obtain a mean close to zero. The review of the literature reveals that normalization could improve performance of the model (Sola & Sevilla, 1997). Normalizing the data generally speeds up learning and leads to faster convergence. Accordingly, mapping data to around 0 gains much faster training speed than mapping them to the intervals far away from 0 or using un-normalized raw data.

Task 4 – ANN-SOM Algorithm

Formally this stage is about selecting an appropriate Machine Learning Algorithm. This is an iterative process. During this study, we identified multiple machine learning algorithms applicable to our data and VAT fraud detection challenges. Regardless, we alluded previously that we shall use an unsupervised learning approach which is appropriate for unlabelled data. The algorithms we evaluated were K-means and Self Organizing Maps (SOM). According to Riveros et al. (2019) the model trained with SOM outperformed the model trained with K-means. In their study they found that the SOM improved detection of patients having vertebral problems (Riveros, Cardenas, & Pico, 2019). Likewise, after a few iterative processes, comparing the SOM and K-means performance, we chose ANN-SOM algorithm.

Task 5 – Train and Test Model

This stage is concerned with creating a model from the data given to it. At this stage we split the dataset into training and test dataset into. Accordingly, we cut the dataset into 20% for testing and 80% for training. Herein, the training process is unsupervised. The remaining dataset is then used to evaluate the model. These two steps are repeated a number of times in order to improve the performance of the model (Riveros, Cardenas, & Pico, 2019).

Task 6 – Optimize Model

A model's first results is not its last. The object of the optimization or tuning we conducted was to improve performance of the model. Tuning a model involves changing hyper parameters such as learning rate or optimizer (Bennett & Parrado-Hernandez, 2006). The purpose for tuning and improving the model is repeatability and efficiency. Someone should be able to reproduce the steps one has taken to improve performance. Additionally, we optimized the model to reduce training time.

Task 7 – Deploy Model

The aim of this stage is the proper functionality of the model after deployment. The models should be deployed in such a way that they can be used for inference as well as they should be updated regularly (Bennett & Parrado-Hernandez, 2006).

Task 8 – VAT audit Case selection

The cohort of VAT vendors with return declarations that have been identified by the SOM as suspicious land up in the “funnel” for further scrutiny. This step is comprised of human verification. This audit is merely a general audit of cases selected for further scrutiny. This contrasts with Investigative Audit, which is concerned with the auditing of cases by a specialist auditor.

Task 9 –Investigative Audit, Criminal Investigation and Enforcement

Investigative audits are different from other tax audits in that a centralised specialist team conducts them. Task 9 is undertaken based on the results obtained from the previous audits conducted in Task 8 above, where audit officers have identified evidence of serious fraud.

Task 10 - Tax Compliance

The tax compliance task is involved with the scrutiny of compliance related attributes like filing returns on time, timely payments, accurate completion of returns and timely registration with the tax authority, among others.

Task 11 – Voluntary Compliance

The aim of the VAT fraud detection AI framework is to increase voluntary compliance. The level of audit activity and frequency of audit will be dictated by the availability of staff resources. The convenience of the AI framework suggested herein, is that it will ensure that the available staff resources are deployed judiciously with the twin objectives of maximizing both revenue collection and voluntary compliance by VAT dealers.

The “filter” or “funnel” denoted by the number 8, symbolizes the audit process, which involves a detailed human verification and validation of lading. This in turn assists in the independent verification financial records such as sales invoices, purchase invoices, customs documents, and bank cash deposits. However, the scope of the human verification is limited to the subset of taxpayers that have been flagged as anomalies by the SOM algorithm we propose. Once human verification has confirmed the presence of suspicious VAT declarations, such cases are then dealt with in Step 9. Step 9 is a depiction of the work performed by investigative audit, criminal investigation, and enforcement teams, on confirmed cases. With this framework we envisage, that the effectiveness and efficiency of this AI assisted compliance framework will enhance detection of suspicious VAT vendors. Consequently, we anticipate that tax compliance will improve as the fear of detection increases (Step 10). Voluntary compliance will be a consequence of an improved, effective, and efficient AI based case selection technique (Step 11).

4.1 Data Collection

We employ a rich data set, that is, the totality of VAT returns covering 6 years from 2013 to 2018. In order to delineate our data collection techniques, we chose to concentrate on only one type of industry that is mining. We have collected VAT returns for the complete list of registered vendors for six tax years, that is, 2013 to 2018. While the firms have been anonymised so we cannot link them with any publicly available data, they are assigned identifying numbers so that we can follow a firm over time. The data contains detailed information on the line items in the returns, which is VAT 201 declaration form of the South African tax administration. For instance, from the VAT return we managed to acquire 35 continuous variables. For ethical and confidentiality reasons, we shall not list all 35 variables, but only a subset of variables (Table 1).

4.2 Data Preparation

According to Peck et.al (2015) data preparation is the cleansing and organizing of real-world data, which is known to consume more than 80% of the time of a data scientist’s work. Real-world data or raw data is dirty, full of missing values, duplicates and in some cases incorrect information (Peck, Olsen, & Devore, 2015).

Most machine-learning algorithms cannot deal with missing values. Hence, the data needs to be converted and cleansed. In handling missing values, we dropped rows and then applied linear interpolation using mean values. Depending on the importance of the variable or feature and amount of the missing values, any one of these solutions can be employed (Peck, Olsen, & Devore, 2015).

Regardless, we are fortunate in that we obtained a clean and high-quality dataset. However, the VAT, return dataset we obtained was at monthly level. We subsequently summed up all variables to annual values. The aggregation of all numerical variables of the VAT returns spans a period beginning in 2013 to 2018. Thereafter the rand value amount was converted into ratios for ease of comparison.

Nevertheless, as stated before the details of some of the variables that we used in this study could not be reported herein, due to the confidential nature of the tax audit process. Doing so can increase the potential for reverse engineering of the audit process, which is clearly undesirable and unlawful. However, each VAT ratio is designed from a point of view that a significantly higher or lower ratio value in relation to the rest of the sample or observations could arouse suspicion. In the opinion of Pamela Castellón González and Juan D. Velásquez fraud cases are most likely to occur among the extreme values of variables (González & Velásquez, 2013).

4.3 Dataset

The dataset consists of 5065 observations with 35 continuous variables. The observations are of Value Added Tax declarations or returns filed with the South African tax administration. The dataset consists of variables that have been captured on the South African tax administration VAT return form (VAT 201 form), by the taxpayers. Due to privacy and data confidentiality concerns, the taxpayer personal information details are not presented in this study. For the same reason, the complete data description or variables cannot be shown in this paper. However, the structure of the dataset showing sample variables can be seen in (Table 1).

However, we do provide aggregate indicative results that demonstrate the effectiveness of our approach. In general, the 35 variables or attributes we selected for this study can be categorized into: gross income profile, income source, expense profile, source of purchases, tax payable or refundable, sales destination, imports or export purchases, accuracy of the declarations, overdue payments of taxes due to the tax authority, market segments, taxpayer industry, demographics, and the size of the firm.

Table 1

Structure of the Mining Industry Dataset

Table 2

VAT Dataset Variables Descriptions

4.4 Exploratory Data Analysis

In this section we will use graphs, visualization, and transformation techniques to explore the VAT dataset (Table 2) in a systematic way. Statisticians call this task, exploratory data analysis, or EDA for short. EDA is a repetitive cycle to firstly, give rise to questions about our data. Secondly, during this phase, we look for answers by visualizing and transforming the dataset population and lastly, we use what we have learnt to refine the questions and/or generate new questions. EDA is not a formal process with a strict set of rules (Peck, Olsen, & Devore, 2015). We hope this initial data analysis will provide insight into important characteristics of the data.

Furthermore, we anticipate that it can provide guidance in selecting appropriate methods for further analysis. Additionally, we shall use summary statistics (Table 3) to provide information about our dataset. During this stage, we envisage that the summary statistics will tell us something about the values in the dataset. This includes where the average and median lies and whether our data is skewed or not. According to Peck, Olsen, and Devore (2015), summary statistics fall into three main categories. That is, measures of location, measures of spread and graphs (Peck, Olsen, & Devore, 2015).

The measures of location will tell us where the data is centred. It will also tell us where a trend lies. Therefore, we shall use Mean, Median and Mode. The arithmetic mean, also called the average, is the central value of a discrete set of numbers. Specifically, it is the sum of the values divided by the number of values. The median is the middle of a data set. The mode of the data set tells us which value is the most common. On the other hand, measures of spread will tell us how spread out or varied is our data set. According to (Peck, Olsen, & Devore, 2015), the Range (inclusive of the Interquartile range and the Interdecile range), the Standard deviation, the Variance and the Quartiles are examples of measure of spread. The Range depicts how spread-out, is our data. The Interquartile range will tell us where the middle 50 percent of our data is located. Whilst the Quartiles will illustrate the boundaries of the lowest, middle, and upper quarters of the dataset (Peck, Olsen, & Devore, 2015).

A correlation matrix was used to quantify dependencies between 35 continuous variables. For this, a Pearson correlation matrix was calculated for all 35 variables. A correlation value between two variables that has an absolute value of greater than 0.7 is considered as high and therefore they are highly related to each other (Ratner, 2009). The objective of this analysis is to establish whether two variables are correlated to each other, and not that they are necessarily causal (Table 4). The sign of the actual value, which is either positive or negative, provides information about whether two variables are positively or inversely related to each other (Ratner, 2009). For example, the correlation value between sales and input vat is 0.98, meaning that there is a direct positive relationship between the two variables. This is rational, because input vat is charged on all purchases of goods and services, which will later become sales of goods and services by the entity.

The correlation matrix and heat maps generated across the 35 variables are a valuable visual representation of VAT data set trends. While the correlation matrix and the heat maps obtain the same conclusion, the heat maps can provide further information about the distribution and localization of correlated variables. Our method of generating heat maps is capable of visualizing the correlations between multiple variables, providing a broader analysis than using a correlation matrix.

4.4.1 Summary Statistics

Table 3

Summary Statistics

4.4.2 Correlation Matrix

Table 4

Correlation Matrix

4.4.3 Correlation Plots

4.4.4 Scatter Plot

4.4.5 Bubble Chart

5.1 Data Pre-processing

Firstly, as part of the data pre-processing, we use a technique called normalization or standardization, to rescale our input and output variables prior to training the SOM model. The purpose, of scaling the dataset, when training a SOM model, is to normalize the data to obtain a mean close to zero (Serra, 2003). The review of the literature reveals that normalization could improve performance of the model (Sola & Sevilla, Importance of input data normalization for the application of neural networks to complex industrial problems, 1997). Normalizing the data generally speeds up learning and leads to faster convergence. Accordingly, mapping data to around 0 gains much faster training speed than mapping them to the intervals far away from 0 or using un-normalized raw data. Academic researchers (Sola & Sevilla, Importance of input data normalization for the application of neural networks to complex industrial problems, 1997) point to the importance of data normalization prior to the neural network training to fasten the calculations and obtain satisfactory results in nuclear power plant application. In the opinion of (Jayalakshmi & Santhakumaran, 2011) various statistical normalization techniques enhance the reliability and the performance of the trained model.

5.2 SOM Training Algorithm

In training algorithm, the SOM map is trained iteratively by taking training data vectors one by one from a training data vector sequence, finding the Best Matching Unit (BMU) for the selected training data vector on the map and updating the BMU and its neighbours closer toward the data vector. This process of finding the BMU and updating the prototype vectors are repeated until a predefined number of training iterations or epochs is completed. The SOM training progress plot is depicted in Fig. 8 below.

5.3 SOM Neighbour Distance

The neighbour distance is often referred to as the “U-Matrix”. This visualisation is of the distance between each node and its neighbours. Typically viewed with a grayscale palette, areas of low neighbour distance indicate groups of nodes that are similar. Areas with large distances indicate the nodes are much more dissimilar. Furthermore, they indicate natural boundaries between node clusters. Them SOM Neighbour distance plot (Fig. 9)

5.4 Artificial Neural Network-SOM Heat map

The ANN-SOM heat map is the outcome of the Neural Network Self-Organising map (SOM) algorithm we trained on the VAT dataset, which has 35 continuous numeric variables. The heat map shows the distribution of all variables across the SOM. We stated before that the dataset used in this experiment spans a period of 6 years, beginning from year 2013 to 2018. The outcome is a grid of 16 Nodes from 5065 observations belonging to the mining industry (Fig. 10)

5.5 Interpretation of the results

In section above, our VAT dataset consists of 35 continuous variables. The continuous variables are made up of variables like Output VAT, Input VAT, Sales, Cost of Sales, VAT Refund, VAT Payable, to name just a few. The sample size is made up of 5065 different VAT taxpayers all belonging to the mining industry. In our experiments, the ANN-SOM map size is 4x4. The SOM cluster heat map contains sixteen distinct clusters and the dots on the cluster represents individual taxpayers or entities. In the ANN-SOM Heat map above we observe that VAT vendors or entities, for example with similar VAT return characteristics, are grouped in the same area or node. In business, users are more interested in “abnormal clusters” or hot spots. That is, clusters of VAT vendors who have suspicious behaviour than normal nodes or clusters. We use three approaches to identify hot spots. That is, by using the ANN-SOM Heat map. Distance matrix Visualisation and domain experts’ feedback based on component plane visualizations.

Using distance matrix visualizations, homogenous clusters (low variation) will have shorter neighbour distances (the white regions) compared to high variation clusters (the dark regions) as shown in (Fig. 9) The value of a component in a node is the mean value of entities (VAT Vendors) in the node and its neighbours. The average value of entities is determined by the neighbourhood function and the final radius used in the final training (Fig. 8). The colour coding of the map is created based on the minimum and maximum values of the component of the map. In this research paper, we use the grey colour map where the maximum value is assigned black, and the minimum value is assigned white.

However, when interpreting the ANN-SOM Heat map, the abnormal clusters are those that have a fewer number of entities. That is, these nodes are composed of suspicious VAT vendors. Such VAT vendors require detailed human verification by VAT audit specialists. Node 4, for example has the largest number of entities at 4948. The entities clustered in Node 4 are homogeneous in nature, and thus depicts VAT entities with normal behaviour. VAT fraud or suspicious behaviour can be differentiated by observing VAT declarations form attributes such as VAT Liability, Exempt supplies, Diesel Refund, and Input VAT on Capital Goods purchased.

Detection of suspicions VAT declarations is a very challenging task as VAT declarations dataset are extremely unbalanced in nature. Furthermore, the tax fraud domain is full of unlabeled data, which in turn makes it difficult to use supervised learning approaches. In this research paper, we proposed an unsupervised learning approach that detects tax fraud in real time. This has broadened the understanding and knowledge of the VAT fraud phenomenon among researchers and in government marketing domain. Remarkably supervised learning algorithms have proved to be limited in the arena of VAT fraud detection, since the tax administrations have extremely low to non-existent labelled historic data. This in turn cripple the efficacy of supervised learning approaches.

In as much as this paper’s focal point is on VAT fraud detection, we are confident that the present model may be applicable to other tax types, like Industry Income Tax and Personal Income Taxes. This research outcome shows how artificial intelligence techniques could be able to solve VAT fraud. Furthermore, this review put forward high-level and detailed classification digital frameworks on VAT fraud detection. Additionally, framework proposed herein, present tax auditors with a systemic case selection guide of suspicious VAT returns.

Consequently, combining the two frameworks into a single hybrid approach can improve the success of detecting other VAT fraud schemes. Tax administrations may be able to select the most appropriate unsupervised learning technique from this work having considered other alternatives, their operational requirements and business context. Thus, leading to a multitude of available Artificial Intelligence aided VAT fraud detection algorithms and approaches. Additionally, the techniques proposed in this paper will help tax administrations with precise case selection using an empirical and data-driven approach, which does not depend upon labelled historic VAT d Furthermore, we envisage that our approach will improve high rates on suspicious VAT returns, and thus improve tax compliance due to increased likelihood of detection.

We have demonstrated the use of ANN-SOM in exploring hot spots in a large real-world Value-Added Tax domain. Based on our experiments, our approach is an effective instrument for hot spots and heat map exploration since it employs visualizations techniques that are easy to understand. In future different profiling or clustering algorithms and sampling techniques can be applied to further improve the performance of the approach proposed. Furthermore, hybrid algorithms may also be employ in future research outcomes.

Ethics approval and consent to participate- Ethical Approval was granted by the University of South Africa Ethics Committee, College of Science, Engineering and Technology.
Consent for publication- Not Applicable
Availability of data and materials- Not Applicable
Competing interests- No competing interest- (Not Applicable)
Funding- Not Applicable
Authors Contributions: Equal Contributions
Acknowledgements: Not Applicable
Authors information (optional)- NA

Aleksander, I. (2017). Partners of humans: a realistic assessment of the role of robots in the foreseeable future. Journal of Information Technology, 32 (1), 1-9.
Allingham, M., & Sandmo, A. (1972). Income Tax Evasion:A theoretical analysis. Journal of Public Economics, 1(34), 323-338.
Ambrecht. (1998). Increasing Taxpayers Complince : A discussion of the Negligence Penalty. Retrieved May 29, 2021, from www.taxlawsb.com/resources/estates/wash.htm
American Marketing Association. (2017). Definitions of Marketing. Retrieved May 30, 2021, from https://www.ama.org: https://www.ama.org/the-definition-of-marketing-what-is-marketing/
Araujo, E., Silva, C., & Sampaio, D. (2008). Video Target Tracking by using Competitive Neural Networks. WSEAS Transactions on Signal Processing, 4(8), 420-431.
Baesens, B., Veronique, Vlasselaer, V., & Verbeke, W. (2015). Fraud analytics using descriptive, predictive, and social network techniques: a guide to data science for fraud detection. Hoboken, NJ, USA: John Wiley & Sons.
Bandara, K. G., & Weerasooriya, W. M. (2019). A Conceptual Research Paper on Tax Compliance and Its relationships. International Journal of Business and Management, 14(10).
Bankole, F.O., & Bankole, O.O (2017). The effect of Cultural Dimension on ICT Innovation: Empirical Analysis of Mobile Phone Services. Telematics and Informatics, 34 (2017), 490-505
Bansal, M., & Suman. (2014). Credit Card Fraud Detection Using Self Organised Map. International Journal of Information & Computation Technology, 4(13).
Bendor-Samuel, P. (2018). What is a digital platform? The Enterprisers Project. A community helping CIOs and IT leaders solve problems. https://enterprisersproject.com/article/2018/12/what-digital-platform
Bennett, K. P., & Parrado-Hernandez, E. (2006). The Interplay of Optimization and Machine Learning Research. Journal of Machine Learning Research, 7, 1265-1281.
Bergman, M., & Nevarez, A. (2006). Do Audits Enhance Compliance? An Empirical Assessment of VAT Enforcement. National Tax Association, 59(4), 817-832.
Bird, R. M., & Milka, C. D. (1992). Improving Tax Administration in Developing Countries. Washington, D.C: International Monetary Fund.
Dignan, L. (2018, September 2). The AI, machine learning, and data science conundrum: Who will manage the algorithms? Artificial Intelligence. Retrieved May 30, 2021, from https://www.zdnet.com/article/the-ai-machine-learning-and-data-science-conundrum-who-will-manage-the-algorithms/
Feld, L., Frey, B., & Targler, B. (2006). Rewarding Honest Taxpayers ? Paper presented on April 9-11, 2006 on “Managing and Maintaining Compliance“.
Flach, P. (2012). Machine Learning :The Art and Science of Algorithms that Make Sense of Data. University of Bristol, United Kingdom: Cambridge University Press.
Furnharn, T. (1983). The Protestant Work Ethic, Human Values and Attitudes towards Taxation. Journal of Economic Psychology, 3(2), 113-128. doi:http://dx.doi.org/10.1016/0167-4870(83)90068-5
González, P. C., & Velásquez, J. D. (2013). Characterization and detection of taxpayers with false invoices using data mining techniques. Expert Systems with Applications, 40(5), 1427-1436.
Georgieva, Markova, & Pavlov. (2019). Using Neural Network for Credit Card Fraud Detection. AIP Conference Proceedings, 2159. doi:https://doi.org/10.1063/1.5127478
Hagiu, A. (2009). Two-sided platforms: Product variety and pricing structures. Journal of Economics and Management Strategy, 18, 1011–1043
Haselton, T. (2018, May 8). Google’s Assistant is getting so smart it can place phone calls and humans think it’s real. Retrieved from https://www.cnbc.com: https://www.cnbc.com/2018/05/08/googles-assistant-will-soon-place-phone-calls-to-book-appointments.html
Huete-Alcocer, N. (2017) A Literature Review of Word of Mouth and Electronic Word of Mouth: Implications for Consumer Behavior. Front. Psychol. 8:1256. doi: 10.3389/fpsyg.2017.01256
Jayalakshmi, T., & Santhakumaran, A. (2011). Statistical normalization and back propagation for classification. International Journal of Computer Theory and Engineering, 3(1), 89-93.
Keen, M., & Smith, S. (2007). VAT Fraud and Evasion: What do we know, and what can be done ? International Monetary Fund, Fiscal Affairs Department. International Monetary Fund.
Kelechi. (2018). Retrieved from https://medium.com/coinmonks/handling-imbalanced-datasets-predicting-credit-card-fraud-544f5e74e0fd
Kirchler, E. (2007). The economic psychology of tax behaviour. Cambridge, United Kingdom: Cambridge University Press.
Kohonen, T. (1982). Biological Cybernetics, 43, 59-69.
Krishnamurthya, R., & Desouza, K. C. (2014). Big data analytics: The case of the social security administration. Information Polity, 19(1), 165-178.
Larsson, O (2021). AI & Digital Platforms: The Markets. In AI and Learning Systems-Industrial Application and Future Direction. IntechOpen Book Series. DOI: 10.5772/intechopen.93098.
McIntyre, D. P., & Srinivasan, A. (2017). Networks, platforms, and strategy: Emerging views and next steps. Strategic Management Journal, 38, 141–160.
Madongo, I. (2017). What is VAT fraud and why is the EU worried? KYC360 News. Retrieved Mar 10, 2021, from www.riskscreen.com: https://www.riskscreen.com/kyc360/article/vat-fraud-eu-laws/
Mahadevan, B. (2000). Business Models for Internet based E-Commerce: An Anatomy. Carlifornia Management Review, 42(4), 55-69.
Marketing Evolution. (2021). https://www.marketingevolution.com/marketing-essentials/ai-markeitng. Retrieved May 29, 2021, from marketingevolution.com.
Molsa, M. (2017). Success factors when implementing AI-powered marketing solutions.
Murphy, K. (2004). An examination of taxpayers’ attitudes towards the Australian tax system: findings from a survey of tax scheme investors. Australian National University, Research School of Social Sciences,. Canberra, Australia: Centre for Tax System Integrity, Working Paper No. 46.
Nagadevara, V., & Kumar, A. (2006). Development of Hybrid Classification Methodology for Mining Skewed Data Sets – A Case Study of Indian Customs Data. Proceedings of the 4th ACS/IEEE International Conference on Computer Systems and Applications. United Arab Emirates.
Peck, R., Olsen, C., & Devore, a. L. (2015). Introduction to Statistics and Data Analysis (5th ed.). Brokes Cole.
Ratner, B. (2009). The correlation coefficient: Its values range between + 1 / - 1, or do they ? International Journal of Computer Theory and Engineering, 17, 139-142.
Riveros, N. A., Cardenas, B. A., & Pico, E. E. (2019). Comparison between K-means and Self-Organizing Maps algorithms used fordiagnosis spinal column patient. Informatics in Medicine Unlocked, 16. doi:https://doi.org/10.1016/j.imu.2019.100206
Serra, P. (2003). Measuring Performance of Chile's Tax Administration. National Tax Journal, 56(2), 373-378.
Shao, H., Hong, Z., & Chang, G.-R. (2002). Applying Data Mining to Detect Fraud Behavior in Customs Declaration. Proceedings of the InternationalConference on Machine Learning and Cybernetics, 3,
Siddharth, M., Hao, L., & Jiabo, H. (2019). Robust geomechanical characterization by analyzing the performance of shallow-learning regression methods using unsupervised clustering methods.10.1016/B978-0-12-817736-5.00005-3. Houston, Texas: Gulf Professional Publishing.
Silvani, C. (1992). Improving Tax Compliance. Washington, D.C, USA: In Improving Tax Administration in Developing Countries, Richard Bird and Milka Casanegra de Jantscher.
Slemrod, J. (1992). Do Taxes Matter ? Lessons from the 1980's. American Economic Review, 82(2), 250-256.
Smith, W. H (1970). Information Systems in Tax Administration. In proceedings of the Annual conference on Taxation under the Auspices of the National Tax Association. Vol 63 (1970), pp 267-283.
Smola, A., & Vishwanathan, S. (2008). Introduction to Machine Learning. Cambridge, UK: Cambridge University Press.
Sola, J., & Sevilla, J. (1997). Importance of input data normalization for the application of neural networks to complex industrial problems. IEEE Transactions on Nuclear Science, 44(3), 1464-1468.
Sola, J., & Sevilla, J. (1997). Importance of input data normalization for the application of neural networks to complex industrial problems. IEEE Transactions on Nuclear Science, 44(3), 1464-1468.
Stallkamp, M., & Schotter, A. P. J (2021). Platforms without borders? The international strategies of digital Platform firms. Global Strategy Journal, 11(1), 58-80. doi/epdf/10.1002/gsj.1336.
Swarnajyoti, B., & & Lorenzo, P. (2014). A novel SOM-SVM based active learning technique for remote sensing image classification. IEEE Transaction on Geoscience and Remote Sensing, 52(11), 6899-6910. doi:DOI: 10.1109/TGRS.2014.2305516
Tripathi, A. (2017). Practical Machine Learning - Cookbook. Birmingham, UK: Packt Publishing Ltd.
Vanhoeyveld, J., Martens, D., & Peeters, B. (2019, October 31). Value-added tax fraud detection with scalable anomaly detection techniques. Applied Soft Computing Journal, 86, 1-20. Retrieved June 07, 2020
Weisman, J. (2001, February 16). IRS audits at record low: Officials cite decade-long staff decrease, old computers,law defending taxpayers. USA Today.
White, M. J. (2014, Jan 27). The SEC in 2004. 41st Annual Securities Regulation Institute. Coronado, California. Retrieved from https://www.sec.gov/news/speech/2014-spch012714mjw

Tables 1 to 4 are available in the Supplementary Files section

No competing interests reported.

Tab1.png
Table 1 - Structure of the Mining Industry Dataset
Tab2.png
Table 2 – VAT Dataset Variables Descriptions
Tab3.png
Table 3 - Summary Statistics
Tab4.png
Table 4 - Correlation Matrix

Download PDF

Version 1

posted

You are reading this latest preprint version

Artificial Intelligence Systems for Value Added Tax Collection via Self Organizing Map (SOM)

Status:

Version 1

Abstract

Figures

1. Introduction

2. Background

2.1 Context

2.2 Theoretical background

2.2.1 Artificial Intelligence (AI) Digital Platform

2.2.2 Social Behaviours on Tax Fraud and Compliance

3. Vat Fraud Detection Ai Framework

3.1 Overview of the Framework

3.2 Flow of the Framework

4. Data Analysis

4.1 Data Collection

4.2 Data Preparation

4.3 Dataset

4.4 Exploratory Data Analysis

4.4.1 Summary Statistics

4.4.2 Correlation Matrix

4.4.3 Correlation Plots

4.4.4 Scatter Plot

4.4.5 Bubble Chart

5. Experiments

5.1 Data Pre-processing

5.2 SOM Training Algorithm

5.3 SOM Neighbour Distance

5.4 Artificial Neural Network-SOM Heat map

5.5 Interpretation of the results

6. Discussion And Conclusion

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1