Keywords

Money Laundering, Anti-Money Laundering Recognition, Gradient Boosting Classifier, Cash Rules

Introduction

One of the most alarming threats to the stability and progress of the economy of the world is the monster called money laundering (Schneider & Windischbauer, 2008). To prevent this monster and the gruesome damages it causes, certain guidelines are put in place. These guidelines are called Anti–Money guidelines. For a smooth strategy regarding money laundering, it has been recommended that financial institutions move away from approaches that are rule-oriented and risky, to those that present risk as their main drivers (Savona and Riccardi, 2019). Not too long ago, the Fifth Anti-Money Laundering Directive (5AMLD), was adopted by the European Union. This move by the European Union is to amongst other things, fight the sponsoring or financing of terrorism and money laundering. This law applies to entities that are directly dealing with users. It ensures that there is a proper application of customer due diligence, transactions monitoring, and customers’ history maintenance and lastly to ensure that every transaction is reported especially those appearing suspiciously linked to terrorists or money launders. It goes further to clearly state that this law applies basically to exchanges involving crypto, fiat, and wallets.

Since financial crimes can be committed irrespective of location, guidelines of the same sort have been recently set up to monitor and regulate the cryptocurrency markets in the United Kingdom and the United States. It was discovered that financial criminals have a way of injecting funds gained illicitly back to the financial system with a very low chance of being detected by the authorities, and so these guidelines were set to prevent such in the financial market. They are made very compulsory and as such heavy fines are placed for financial institutions with a credible Anti-money Laundering directive. Several institutions have been involved in such a one worthy of note is the Danske Bank Scandal (Yicheng, 2015). Emerging trends within the financial system have recently gained wide popularity and embrace across climates and regions and with these developments, come the various concepts that if not properly checked and regulated, will aid money laundering activities (Bynagari, 2015). One of such popular trends in cryptocurrency; this technology is pseudonymous in nature and could be a breeding ground for criminals to hide behind and perpetuate money laundering. Regulations have been established to curb this menace of new technologies aiding money laundering by shielding the criminals behind the act (Lessambo, 2020).

Statement of Problem

There has been an old method of detecting money laundering and this method is famously regarded as rule-based. This system consists of a set of conditions that check whether certain thresholds are exceeded or events occur (Vadlamudi et al., 2021). That is to say, that the entire activity is monitored closely with eyes on the limit of activity such as volume of money or destination, and once an unlikelihood is detected, it becomes a suspicious signal and is reported and investigated. This technique has been very effective and has been around for ages, however, its outstanding shortcoming is that it usually could lead to a high false-positive rate and require experts in the domain to create rules with which to decide which acts are laundering and which may not be (Jullum et al., 2020). To overcome these setbacks, machine learning is employed, and the machine learning algorithm does this by inferring patterns from previous data. This inference can decrease the rate of this false positive rate while keeping the false-negative rate, reduced to the barest minimum, as low as possible, if not cut off.

Existing literature has identified areas that can be explored further for more understanding. Various ensembles could be used, but tree-oriented ensembles are one of the most used. First of all, the Random Forest and the tree-oriented gradient boosting (Baek, et al., 2019) and (Farrugia, et al., 2020). Of the two ensembles, it is not clear which of them is more effective for detecting fraud activities, especially in cryptocurrency networks. As evident in available researches, most of the works are based on the detection of such fraud activities either the account or transactional levels (Lin, et al., 2019).

In an attempt to compare the random forest and the tree-based ensemble against each other, at the account level, the RF performed better than the tree-based. However, some researchers argue against it (Sun et al., 2019; Toyoda et al., 2017).

Aim and Objectives of the Study

This study aims to investigate the performance of a gradient boosting algorithm in identifying money laundering activities.

To achieve the aim of this study, we will attempt to do the following;

Identify the performance of an offline gradient boosting algorithm in identifying genuine and fraudulent activities at the account level and transactional levels in the financial system.

Use selected data sampling techniques to improve the detection of both genuine and fraudulent activities at a transactional level using a gradient boosting algorithm.

Literature Review

Scholars have reported in their publications have claimed that tree-based classifiers can be employed to effectively detect money laundering activities (Harlev et al., 2018); Jullum et al., 2020) attempted to detect laundering at a transactional level, using the XGBoost. The XGBoost was employed due to some attributes it possesses such as efficiency, scalability, and its ability to reduce time. It was observed that the XGBoost, performed better than a rule-based system. Hitherto, Senator et al. (1995), showed that the tree-based models could be useful in money laundering detection especially in traditional finance. But because of its limitation in lacking labeled data, it is difficult to use the methods. Machine learning has been employed in the detection of money laundering and related financial crimes, especially in crypto-currency. The money laundering process is equivalent to the traditional process used in traditional everyday finance.

According to Khan et al. (2021), to effectively classify, whether transactions are associated with money laundering or not, the random forest can be used together with network analysis and community detection. We will attempt to illustrate with a simplified diagram, the various types of machine learning approaches that could be employed, these techniques include; supervised, semi-supervised and unsupervised. Baek, et al., (2019) investigated the labeled accounts, various unsupervised approaches including k-means and kd-trees. This approach was aimed at reducing the effect of unlabeled data indicating financial crimes.

By employing heuristic-based reasoning processes, to categorize instances, researchers have paid attention to detecting a specific kind of criminal activity (Ahmed & Ganapathy, 2021). Such criminal activities as accounts linked to Bitcoin and Ponzi schemes (Bartoletti et al., 2018). Others link accounts or transactions to multiple labels in what is known as multiclass classification. Under this classification, some activities are mixer services, dark marketplaces, exchanges, wallet providers, scams, and others such as gambling, terrorism even ransomware (Zola et al., 2019). Binary classification and grouping of activities based on activities. These activities are grouped into illegal and illegal (Lee et al., 2020). To combat money laundering effectively, it is necessary to classify accounts to their service and this classification comes from various areas in the network. However, not the entire network is exploited, rather, dark marketplaces, gambling sites, and mixing services (Ahmed et al., 2021). A careful glance at previous researches along these lines indicates that the techniques employed have either used supervised learning or unsupervised learning. The unsupervised learning labels the data and uses a supervised learner to conclude the label (Farrugia et al., 2020). However, these classifications have shown weakness on two levels; first of all on the accounts level and then on the transaction level (Lee et al., 2020). Researchers have investigated several machine learning models for both supervised and unsupervised learning. A few examples of these models are; Extra Trees (ET), AdaBoost, Bayes Network (BN), Logistic Regression (LR), Multi-layer Perceptron (MLP), etc., (Harlev et al., 2018). Of all of the prototypes, the most common are Random Forest models and gradient boosting (Zola et al., 2019; Ganapathy, 2021). There has been a debate as to which model performs better; in some cases, the Random Forest models did perform better than the gradient boosting, and some more deep learning models like the Graph Convolution Network as investigated by (Toyoda et al., 2017; Weber et al., 2019). Although these two ensembles have never been compared side by side on both levels of accounts and transactions, the gradient boosting algorithm has been seen to outperform the Random Forest prototypes (Lin et al., 2019). To use the gradient boosting algorithm in detecting money laundering, one will need to consider the datasets in which cases are heavily unbalanced. To tackle this, we will need to utilize techniques that are cost-effective and at the same time are flexible in their data sampling approaches (Sun et al., 2019).

Methods

Two frameworks were employed to achieve the objective of this study. The chosen frameworks consist, the module works out / test division analysis that is for offline learning as well as prudential analysis that is for online learning. In the workout/test divide analysis, offline prototypes are constructed statistically as they are worked out on a fixed data divide to work out and test groups. On the other hand, the prequential analysis consented to analyze online learners by simulating an unbounded data stream, and it trains accordingly,

1. Divide the existing dataset into groups (according to time steps|)
2. Work out a prototype on timestamp t;
3. Analysis of the prototype on timestep t + 1;
4. Relate steps ii and iii to successive time steps (Hidalgo et al., 2019).

Table 1 presents the datasets utilized along with the frameworks and experiment(s) that were employed, ensured by an overview of the respective experiment.

Table 1  Dataset Utilized Along With The Frameworks And Experiments
Feature	Elliptic	Ethereum	NOAA
N instances	46,560	4680	18,158
N feature	93 LF, 193 LF_NE, 165 AF, 265 AF_NE	41	7
Group ratio	˜1:8	˜1:1	˜1:2
Time steps	48	N/A	605
Experiment 1	Yes	Yes	Yes
Experiment 2	Yes	Yes	Yes
Experiment 3	Yes	Yes	Yes

Experiment 1 (offline context): We postulate that, when contrasted to Random Forest, DT-oriented gradient boosting can enhance the categorization of legal and illegal operations at either the transaction and account levels. We tested yet if the hypothesized offline approaches outperformed Random Forest (the highest performing paradigm in the Weber et al. (2019) analysis). To replicate the outcomes published by Ahmed et al. (2013), the indicated classifiers were applied to the Ellipsoidal data. The preceding division was employed in this offline context:

1. Ellipsoidal, time step ≤ 34 has been utilized for working out, with the preceding time steps for screening;
2. Cryptocurrency illegal Accounts, the first 3,275 examples were utilized for working out, with the surviving incidents being utilized for diagnostics (stratified selection on class);
3. NOAA, time step 423 was utilized for working, with the very next time steps being utilized for experimentation.

To avoid breaking the order of time, data comprising temporal information was divided according to time steps (Bynagari, 2016; Jullum et al., 2020).

Experiment 2 (offline tuning with data-sampling): We assume that DT-oriented gradient boosting combined with data-sampling can boost legal or illegal classification at the process level even further. From this point on, the primary focus changed to transaction detection and measurement because it allows for real-time analysis. We assumed if the data-sampling procedures that were tested decreased the utilizing the same false-negative frequency (through recall assessment) experiment 1: work out/test division.

Experiment 3 (in an online environment): We assume that in an evolving data stream, the suggested ASXGB can enhance the classification of licit/illicit transactions environment. We compared our proposed approach to the following: Prequential analysis is used by other responsive learners. Data sampling was not used due to time restrictions.

On-time steps ≥5 and ≥35, the outcome was excellent. On-time steps ≥5 and ≥35, the efficiency was excellent (consistent with earlier studies) were noted for Ellipsoidal data is a term used to describe a set of data.

Evaluation

All work was carried out on Microsoft Azure Cloud Computing utilizing an H16m instance with the following specs: 16.0 core Intel Xeon E5-2667 v3 Haswell 3.20 GHz with 224 GB DDR4 RAM, Python 3.7.6, and Ubuntu 18.04 LTS (operating system).

Benchmark Prototypes

Various benchmarks were evaluated alongside the proposed offline gradient boosting algorithms and our proposed online model to compare the two algorithms. Random Forest classifier was utilized as a benchmark in the train/test split evaluation and executed utilized Scikit learn to enclose (Pedregosa et al., 2011). The following were the reasons for supporting Random Forest:

1. It outperformed all the methodologies assessed in the original study that given the Euclidean data when compared to other prototypes (Monamo et al., 2016; Ahmed, 2020; Bartoletti et al., 2018; Weber et al., 2019).
2. It outperformed all the approaches evaluated in the original study that offered the Euclidean data when compared to other brands, it proved to be the best performing in similar issues (Monamo, et al., 2016; Bartoletti, et al., 2018; Weber et al., 2019).
3. It outperformed all the methods evaluated in the original study that provided Deep learning is one of them (Weber et al., 2019).

In the comprehensive screening, two online standards, namely the Adaptive Random Forest (ARF) and Adaptive eXtreme Gradient Boosting were reviewed to assess the outcomes obtained by our proposed ASXGB (AXGB). ARF is a variant of random forest that has been tweaked to cope with changing data streams (Gomes et al., 2017; Bynagari, 2019). ARF has shown to be a successful online model (Boiko et al, 2019; Manojkumar et al., 2021), and assuming ASXGB's performance, is a continuation of gradient boosting. ARF is an effective online model (Boiko et al, 2019; Ahmed, 2021), and given that ASXGB is a gradient boosting extension for online learning, it is a natural fit. The RF extension was put to the test. The AXGB is a web-based version of XGBoost (Bynagari, 2016), which spurred the creation of the AXGB. As a result, benchmarking against ASXGB was critical for this type of model. Two strategies for updating the ensemble to tackle idea drift were suggested in the original study (Bynagari, 2014):

1. Push strategy: older prototypes are eliminated before attaching newer models, comparable to First in First Out;
2. Replace strategy: older models are substituted with fresh systems.

Both of these modifications were put to the test and were given the names “AXGB (push)” and ‘AXGB[R]’ (resistance)/ (substitute). Sklearnmulti-flow (Montiel et al., 2018) was used to create the ARF classifier. The default settings needed code for the AXGB, as well as GitHub6, was used to get the hyper factors required for every update modification, which was supplied by (Sharma et al., 2021).

Datasets

This study utilized elliptic data for the transactional level detection because it makes up one of the largest publicly available labeled data in the dynamic financial market, as a follow-up to the work of Weber et al. (2019). This method made it possible to be able to compare our proposed method with previous studies. The elliptic data are in the form of a transaction graph with each transaction forming a vertex. Each transaction is a vertex and the edges that are directed represent the flow of payments, outgoing and incoming (Bynagari, 2014). The labels represent whether the entity carrying out the transaction is authentic that is licit, such as wallets providers, miners, etc, or illicit such as scammers, terrorists, malware, etc.

Every transaction is connected to a time step. These time steps are a representation of the time of confirmation of the transaction and are comprised of a single connected component of transactions that were settled within 3 h or less from one another (Weber et al., 2019).

For the study, the authors used 49-time steps equally distributed within approximately 2 weeks. These steps can be expanded through extrapolation to cover about 98 weeks. The exact period, from start to finish date covered by the datasets was not specified.

The dataset consists of 203,795 transactions with a total of 234,354 directed edges. From all these transactions, 4544 transactions were marked as illicit and 42,018 marked as licit with the remaining being undefined. Every transaction is made up of 166 attributes, categorized as local and aggregated features. The first 93 attributes (local features) include information such as transaction fee, number of outputs/inputs, and time step. The other 71 attributes included aggregated information from one-hop backward forward from the central vertex, such as the standard deviation and correlation coefficients of neighboring transactions, for the same information extracted for local features (Bynagari, 2017). Weber et al. (2019) noted that the models’ performance degraded after time step 42 due to the closure of a dark marketplace (that is a sharp reduction in the number of illicit transactions).

This indicated that the data was nonstationary, and we confirmed this by applying various time plot tests. The unit root test is termed the Augmented Dickey-Fuller test (ADF) (MacKinnon, 1995). In the study conducted by Farrugia et al. (2020), used the Ethereum illegal Accounts data which consisted of 2178 illegal and 2501 legal accounts. These illicit accounts were flagged by the Euthereum community. Unlike the first transaction-based dataset, this second dataset was used to investigate unlawful activities at an account level. These data have a total of 41 features, which were based on the transaction history for specific accounts for instance total ether balance, the difference between the first and last transaction in minutes, and minimum value of ether ever sent). The illicit category in these data represent accounts linked to illicit activities such as scam lotteries, fake initial coin offering, and Ponzi schemes (Farrugia & Azzopardi, 2020).

Results and Discussion

Test Results

The accuracy, sensitivity, recollect, F1-score, non-linear data, and other metrics were used to assess the described classifiers. F1-score over time steps, box plots (when temporal information is available). Because of the nature of the classifiers, the assessment proceeded through hundred repetitions for the reason that it was non-deterministic. The findings were then aggregated. Tables 2, 3, and 4 present the outcomes mined from experiment 1 to 3.

Table 2 Work Out / Test Division Analysis Of Offline Supervised Learning Prototypes Experiment 1
Experiment 1	Accuracy	Sensitivity	Recollect	F1
Prototype	Elliptic dataset
XGBAF	0.976-0.977	0.901-0.920	0.722-0.736	0.802-0.814
XGBA_NE	0.978-0.978	0.978-0.985	0.692-0.691	0.811-0.812
LGBAAF	0.978–0.978	0.930–0.931	0.722–0.731	0.813–0.819
LGBAAF_NE	0.978–0.978	0.982–0.984	0.688–0.694	0.809–0.814
CATAF	0.978–0.978	0.948–0.935	0.720–0.727	0.819–0.818
CATAF_NE	0.978–0.978	0.982–0.974	0.695–0.690	0.814–0.808
RFAF	0.976	0.896	0.720	0.799
RFAF_NE	0.978	0.957	0.714	0.818
Prototype	Ethereum Illegal Accounts dataset
XGB	0.980–0.988	0.988–0.984	.9068–0.980	0.978–0.982
LGBA	0.977–0.980	0.983–0.982	0.967–0.976	0.975–0.979
CAT	0.979–0.979	0.979–0.980	0.976–0.976	0.978–0.978
RF	0.972–0.973	0.981–0.982	0.958–0.960	0.969–0.971
Prototype	NOAA Dataset
XGB	0.775–0.784	0.728–0.748	0.595–0.0.593	0.663–0.654
LGBA	0.783–0.794	0.734–0.749	0.616–0.638	0.670–0.689
CAT	0.791–0.789	0.747–0.743	0.629–0.628	0.683–0.681
RF	0.779–0.769	0.741–0.747	0.588–0.594	0.656–0.661

Table 3 Work Out/Test Division Analysis Of Offline Supervised Learning With Data Sampling In Experiment 2
Experiment 2	Accuracy	Sensitivity	Recollect	F1
Prototype	Elliptic dataset sampled utilizing NCL
XGBAF	0.976-0.977	0.901-0.920	0.722-0.736	0.802-0.814
XGBA_NE	0.978-0.978	0.978-0.985	0.692-0.691	0.811-0.812
LGBAAF	0.978–0.978	0.930–0.931	0.722–0.731	0.813–0.819
LGBAAF_NE	0.978–0.978	0.982–0.984	0.688–0.694	0.809–0.814
RFAF	0.976-0.976	0.939-0.944	0.724-0.725	0.819-0.820
RFAF_NE	0.980-0.980	0.972-0.976	0.720-0.722	0.827-0.822
Prototype	Elliptic dataset sampled utilizing SMOTE
XGBAF	0.976-0.977	0.901-0.920	0.722-0.736	0.802-0.814
XGBA_NE	0.978-0.978	0.978-0.985	0.692-0.691	0.811-0.812
LGBAAF	0.978–0.978	0.930–0.931	0.722–0.731	0.813–0.819
LGBAAF_NE	0.978–0.978	0.982–0.984	0.688–0.694	0.809–0.814
RFAF	0.976-0.976	0.939-0.944	0.724-0.725	0.819-0.820
RFAF_NE	0.980-0.980	0.972-0.976	0.720-0.722	0.827-0.822
Prototype	Elliptic dataset sampled utilizing NCL_SMOTE
XGBAF	0.976-0.977	0.901-0.920	0.722-0.736	0.976-0.977
XGBA_NE	0.978-0.978	0.978-0.985	0.692-0.691	0.978-0.978
LGBAAF	0.978–0.978	0.930–0.931	0.722–0.731	0.978–0.978
LGBAAF_NE	0.978–0.978	0.982–0.984	0.688–0.694	0.978–0.978
RFAF	0.976-0.976	0.939-0.944	0.724-0.725	0.976-0.976
RFAF_NE	0.980-0.980	0.972-0.976	0.720-0.722	0.980-0.980
Prototype	NOAA Dataset Sampled utilized NCL
XGB	0.742–0.745	0.601–0.605	0.826–0.828	0.696–0.699
LGBA	0.732-0.751	0.586-0.612	0.853-0.823	0.695-0.702
RF	0.745-0.745	0.609-0.608	0.798-0.802	0.695-0.702
Prototype	NOAA Dataset Sampled Utilized NCL_SMOTE
XGB	0.778-0.780	0.705-0.710	0.653-0.651	0.678-0.679
LGBA	0.775-0.777	0.683-0.707	0.689-0.645	0.686-0.674
RF	0.770-0.770	0.678-0.676	0.678-0.687	0.678-0.681
Prototype	NOAA Dataset Sampled Utilized NCL_SMOTE
XGB	0.752-0.730	0.619-0.584	0.794-0.842	0.678-0.690
LGBA	0.735-0.735	0.592-0.592	0.831-0.833	0.691-0.692
RF	0.733-0.732	0.592-0.590	0.815-0.819	0.686-0.686

Offline Supervised Tree?Oriented Ensembles

The Light gradient Boosting Algorithm (LGBA) and XGBoost outflanked the Random Forest algorithm in distinguishing illegal exercises both at an exchange and account level, as appeared in Table 2. Measurable importance was identified when applying the Wilcoxon signed-rank test (α = 0.05) in terms of review, exactness, and F1-Score. The CatBoost algorithm did not result on standard with the other gradient boosting prototypes, and so, was dropped in ensuing tests. The utilization of both neighborhood and amassed highlights on the Euclidean data conveyed better outcomes (online with Weber et al. (2019), and so, the LF and LF_NE were presented from Table 3. The outcomes gotten when duplicating the Random Forest algorithms were higher than those detailed in the unique deliberation Weber et al. (2019); the differences between the best scores for both the suggested offline angle boosting models, and the Random Forest, overall highlights sets were: 0% (exactness), +3% (accuracy), +1.1% (review), and +0.1% (F1-Score). These values were marginally higher in comparison to the outcomes gotten by Weber et al. (2019). At an account level, all the recommended offline models outflanked Random Forest. XGBoost was the foremost significant, with a difference of: +1% (precision), +0.6% (accuracy), +2% (review), and +1.1% (F1-Score). Comparing the most elevated F1-Score detailed (0.960) by the XGBoost usage in Farrugia et al. (2020) consider, to the proposed XGBoost, light gradient boosting algorithm, and CatBoost, we have gotten an advancement of +2.3%, +2%, and +1.9%, respectively. This change may be credited to the number of hype factors tuned and the strategy linked to do so. Xia et al. (2017) state that having an endless parameter look space to tune an XGBoost present in conjunction with a productive calculation can progress execution, and so, this considers enhanced ten hyper factors in comparison to the three upgraded by Farrugia et al. (2020). The outcomes gotten from the NOAA dataset strengthened the idea that the suggested offline calculations outflank random forest, as the results generated from this study were comparable to the other datasets in Table 4.

Table 4 Prequential Assessment Of Online Supervised Learning Approaches Utilizing The Elliptic And Noaa Datasets In Experiment 3
Experiment 3	Accuracy	Sensitivity	Recollect	F1	Accuracy	Sensitivity	Recollect	F1
Prototype	Elliptic dataset reporting when time step =5				Elliptic dataset reporting when time step =35
AXGB[R]AF	0.946	0.812	0.679	0.739	0.948	0.589	0.718	0.647
AXGB[R]AF_NE	0.944	0.870	0.687	0.768	0.961	0.712	0.693	0.703
AXGB[P]AF	0.947	0.777	0.737	0.756	0.946	0.571	0.721	0.638
AXGB[P]AF_NE	0.951	0.791	0.764	0.777	0.952	0.673	0.767	0.718
ASXGBAF	0.960	0.827	0.816	0.821	0.957	0.662	0.727	0.788
ASXGBAF_NE	0.959	0.812	0.830	0.821	0.957	0.662	0.727	0.693
ARFAF	0.968	0.985	0.731	0.839	0.976	0.987	0.656	0.788
ARFAF_NE	0.967	0.978	0.723	0.822	0.976	0.986	0.647	0.782
Prototype	NOAA dataset reporting if the time step = 25
AXGB[P]	0.776	0.689	0.527	0.597
ASXGB	0.695	0.515	0.569	0.541
ARF	0.779	0.712	0.500	0.588

Conclusion

This study explored the likely application of choice tree-based eXtreme boosting calculations, related to proficient hyper factor improvement furthermore, information testing methods. A transformation of Outrageous eXtreme Boosting (XGBoost) to deal with advancing datasets (idea float), with the use of summed up stacking to refresh the fundamental group (beforehand constructed students which are done adding to the in the general forecast) was additionally proposed and demonstrated to be successful. Further work should be directed to investigate how streamlining and information testing methods might have been coordinated into our proposed ASXGB as the two strategies further developed execution in a disconnected setting. Future work will likewise mull over potential memory issues ascribed to this proposed technique, to more readily deal with unbounded streams. A potential arrangement is to trade the meta-student with a straightforward perceptron and supplant base models once afloat is distinguished utilizing ADWIN, like how Troupe of Limited Hoeffding Trees work. All the product created prompting this examination has been open-sourced furthermore, made openly accessible on GitHub, together with the inspected information found on Google Drive.

References

1. Ahmed, A.A.A. (2021). Corporate attributes and disclosure of accounting information: Evidence from the big five banks of China. Journal of Public Affairs, 21(3), e2244.
2. Ahmed, A.A.A. (2021). Perception of the audience of interests on the qualitative characteristics of financial reporting. International Journal of Intellectual Property Management.
3. Ahmed, A.A.A., & Ganapathy, A. (2021). Creation of Automated Content with Embedded Artificial Intelligence: A Study on Learning Management System for Educational Entrepreneurship. Academy of Entrepreneurship Journal, 27(3), 1-10.
4. Ahmed, A.A.A., Siddique, M.N.E.A., & Al Masum, A. (2013). Online library adoption in Bangladesh: an empirical study. In 2013 Fourth International Conference on e-Learning" Best Practices in Management, Design and Development of e-Courses: Standards of Excellence and Creativity" (pp. 216-219). IEEE.
5. Ahmed, A.A.A., Paruchuri, H., Vadlamudi, S., & Ganapathy, A. (2021). Cryptography in Financial Markets: potential channels for future financial stability. Academy of Accounting and Financial Studies Journal, 25(4), 1-9.
6. Baek, H., Oh, J., Kim, C.Y., & Lee, K. (2019). A model for detecting cryptocurrency transactions with discernible purpose. In: 2019 Eleventh International Conference on ubiquitous and future networks (ICUFN), IEEE; 713–17.
7. Bartoletti, M., Pes, B., & Serusi, S. (2018). Data mining for detecting bitcoin ponzi schemes. In 2018 Crypto Valley Conference on Blockchain Technology (CVCBT) (pp. 75-84). IEEE.
8. Boiko, Ferreira L.E., Murilo Gomes, H., Bifet A., & Oliveira L.S. (2019). Adaptive random forests with resampling for imbalanced data streams. International Joint Conference on Neural Networks (IJCNN), IEEE, 1–6.
9. Bynagari, N.B. (2014). Integrated Reasoning Engine for Code Clone Detection. ABC Journal of Advanced Research, 3(2), 143-152.
10. Bynagari, N.B. (2015). Machine Learning and Artificial Intelligence in Online Fake Transaction Alerting. Engineering International, 3(2), 115-126.
11. Bynagari, N.B. (2016). Industrial Application of Internet of Things. Asia Pacific Journal of Energy and Environment, 3(2), 75-82.
12. Bynagari, N.B. (2017). Prediction of Human Population Responses to Toxic Compounds by a Collaborative Competition. Asian Journal of Humanity, Art and Literature, 4(2), 147-156.
13. Bynagari, N.B. (2019). GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. Asian Journal of Applied Science and Engineering, 8, 25-34.
14. Farrugia, S., Ellul, J., & Azzopardi, G. (2020). Detection of illicit accounts over the Ethereum blockchain. Expert Systems with Applications, 150, 113318.
15. Ganapathy, A. (2021). Edge Computing: Utilization of the Internet of Things for Time-Sensitive Data Processing. Asian Business Review, 11(2), 59-66.
16. Gomes, H.M., Bifet, A., Read, J., Barddal, J.P., Enembreck, F., Pfharinger, B., ... & Abdessalem, T. (2017). Adaptive random forests for evolving data stream classification. Machine Learning, 106(9), 1469-1495.
17. Harlev, M.A., Sun Yin, H., Langenheldt, K.C., Mukkamala, R., & Vatrapu, R. (2018). Breaking bad: De-anonymising entity types on the bitcoin blockchain using supervised machine learning. In Proceedings of the 51st Hawaii International Conference on System Sciences.
18. Hidalgo, J.I.G., Maciel, B.I., & Barros, R.S. (2019). Experimenting with prequential variations for data stream learning evaluation. Computational Intelligence, 35(4), 670-692.
19. Jullum, M., Løland, A., Huseby, R.B., Ånonsen, G., & Lorentzen, J. (2020). Detecting money laundering transactions with machine learning. Journal of Money Laundering Control.
20. Khan, W., Ahmed, A.A.A., Vadlamudi, S., Paruchuri, H., & Ganapathy, A. (2021). Machine Moderators in Content Management System Details: Essentials for IoT Entrepreneurs. Academy of Entrepreneurship Journal, 27(3), 1-11.
21. Lee, C., Maharjan, S., Ko, K., & Hong, J.W.K. (2019, December). Toward detecting illegal transactions on bitcoin using machine-learning methods. In International Conference on Blockchain and Trustworthy Systems (pp. 520-533). Springer, Singapore.
22. Lessambo, F.I. (2020). Anti-money laundering laws. In The US banking system (pp. 37-66). Palgrave Macmillan, Cham.
23. Lin, Y.J., Wu, P.W., Hsu, C.H., Tu, I.P., & Liao, S.W. (2019). An evaluation of bitcoin address classification based on transaction history summarization. In 2019 IEEE International Conference on Blockchain and Cryptocurrency (ICBC) (pp. 302-310). IEEE.
24. MacKinnon, J.G. (1994). Approximate asymptotic distribution functions for unit-root and cointegration tests. Journal of Business & Economic Statistics, 12(2), 167-176.
25. Manojkumar, P., Suresh, M., Ayub Ahmed, A.A., Panchal, H., Rajan, C.A., Dheepanchakkravarthy, A., ... & Sadasivuni, K.K. (2021). A novel home automation distributed server management system using Internet of Things. International Journal of Ambient Energy, 1-6.
26. Monamo, P. M., Marivate, V., & Twala, B. (2016). A multifaceted approach to bitcoin fraud detection: Global and local outliers. In 2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA) (pp. 188-194). IEEE.
27. Montiel, J., Read, J., Bifet, A., & Abdessalem, T. (2018). Scikit-multiflow: A multi-output streaming framework. The Journal of Machine Learning Research, 19(1), 2915-2914.
28. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... & Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. the Journal of machine Learning research, 12, 2825-2830.
29. Savona, E.U., & Riccardi, M. (2019). Assessing the risk of money laundering: research challenges and implications for practitioners. European Journal on Criminal Policy and Research, 25(1), 1-4.
30. Schneider, F., & Windischbauer, U. (2008). Money laundering: some facts. European Journal of Law and Economics, 26(3), 387-404.
31. Senator, T.E., Goldberg, H.G., Wooton, J., Cottini, M.A., Khan, A.U., Klinger, C.D., ... & Wong, R.W. (1995, August). The FinCEN Artificial Intelligence System: Identifying Potential Money Laundering from Reports of Large Cash Transactions. In IAAI (pp. 156-170).
32. Sharma, D.K., Chakravarthi, D.S., Shaikh, A.A., Ahmed, A.A.A., Jaiswal, S., & Naved, M. (2021). The aspect of vast data management problem in healthcare sector and implementation of cloud computing technique. Materials Today: Proceedings.
33. Sun Yin, H.H., Langenheldt, K., Harlev, M., Mukkamala, R.R., & Vatrapu, R. (2019). Regulating cryptocurrencies: a supervised machine learning approach to de-anonymizing the bitcoin blockchain. Journal of Management Information Systems, 36(1), 37-73.
34. Toyoda, K., Ohtsuki, T., & Mathiopoulos, P.T. (2017). Identification of high yielding investment programs in bitcoin via transactions pattern analysis. In GLOBECOM 2017-2017 IEEE Global Communications Conference (pp. 1-6). IEEE.
35. Vadlamudi, S., Islam, A., Hossain, S., Ahmed, A.A.A., & Asadullah, A.B.M. (2021). Watermarking Techniques for Royalty Accounts in Content Management Websites for IoT Image Association. Academy of Marketing Studies Journal, 25(4), 1-9.
36. Weber, M., Domeniconi, G., Chen, J., Weidele, D.K.I., Bellei, C., Robinson, T., & Leiserson, C.E. (2019). Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial forensics. arXiv preprint arXiv:1908.02591.
37. Xia, Y., Liu, C., Li, Y., & Liu, N. (2017). A boosted decision tree approach using Bayesian hyper-parameter optimization for credit scoring. Expert Systems with Applications, 78, 225-241.
38. Zola, F., Eguimendia, M., Bruse, J.L., & Urrutia, R.O. (2019). Cascading machine learning to attack bitcoin anonymity. In 2019 IEEE International Conference on Blockchain (Blockchain) (pp. 10-17). IEEE.