Project – Explainable Machine Learning for Detecting Mobile Money Fraud in Low-Resource Financial Environments.

CHAPTER ONE

INTRODUCTION

1.1 Background to the Study

The increasing adoption of digital financial services has transformed how individuals and organisations transfer money, make payments, receive funds, and conduct commercial transactions. Mobile banking applications, mobile money platforms, USSD-based services, agent-assisted payments, and digital wallets have expanded access to financial transactions without requiring customers to visit conventional bank branches for every transaction. These developments are particularly important in developing economies, where mobile-based financial services can support financial inclusion and provide convenient payment options for individuals and small businesses. However, the expansion of digital transactions also creates opportunities for fraudulent activities, including unauthorised transfers, account compromise, social engineering, and manipulation of transaction processes. The World Bank’s Global Findex Database 2021 documents the growing role of digital payments and financial account ownership worldwide, including in developing economies. The Central Bank of Nigeria’s Payments System Vision 2025 likewise identifies the importance of a secure, efficient, and inclusive payments ecosystem for Nigeria’s digital economy. As transaction volumes increase, financial service providers require methods that can identify suspicious activities accurately and promptly while limiting unnecessary interruptions to legitimate customers. Consequently, automated fraud detection has become an important application of data analytics and machine learning in digital financial services.

Mobile money fraud occurs when digital payment services are exploited to obtain funds or financial benefits through unauthorised, deceptive, or otherwise fraudulent activities. Such activities may include account takeover, fraudulent transfers, identity-related deception, suspicious transaction sequences, and the misuse of compromised credentials. Conventional fraud detection approaches often rely on predefined rules, transaction limits, manually selected indicators, and previously identified patterns of suspicious activity. These methods remain useful because they are comparatively understandable and can be implemented with limited computational resources. Nevertheless, static rules may struggle when fraudulent behaviour changes, when several individually ordinary transactions form a suspicious pattern, or when legitimate customers exhibit unusual but harmless behaviour. Bolton and Hand (2002), in their review of statistical fraud detection, explain the importance of analytical methods for identifying fraudulent behaviour in large datasets. Phua et al. (2010) similarly review data-mining approaches to fraud detection and identify the role of computational methods in discovering patterns that may not be readily apparent through manual examination. These contributions establish a foundation for investigating machine-learning techniques that can learn from historical transaction data and identify potentially suspicious activity. However, automated systems must be evaluated carefully because false alarms can inconvenience customers, while missed fraud cases can expose users and financial institutions to financial losses.

Machine learning is a branch of artificial intelligence that enables computational systems to identify patterns in data and use those patterns to make predictions or classifications. In mobile money fraud detection, supervised machine-learning models can learn from labelled historical transactions to distinguish between transactions identified as fraudulent and those considered legitimate. Potential approaches include logistic regression, decision trees, random forests, gradient-boosting algorithms, and neural networks. Unsupervised and anomaly-detection methods may also be useful when labelled fraud examples are scarce or when the goal is to identify unusual transactions that differ from established patterns. Whitrow et al. (2009) investigate transaction aggregation strategies for credit card fraud detection, illustrating the importance of examining transaction behaviour rather than relying exclusively on isolated records. Dal Pozzolo et al. (2015) examine adaptive learning for credit card fraud detection and discuss the difficulties associated with changing fraud patterns and evolving data distributions. Although credit card transactions differ from Nigerian mobile money transactions in some operational respects, these studies provide relevant methodological foundations for developing transaction-based fraud detection systems. For the proposed study, machine learning will be used to build a prototype that classifies transactions according to estimated fraud risk. Its effectiveness will be evaluated using suitable performance metrics rather than assumed from the use of an advanced algorithm alone.

Despite their potential predictive capabilities, some machine-learning models are difficult to interpret because their outputs result from complex interactions among multiple variables. In financial fraud detection, a system that labels a transaction as suspicious without explaining the factors behind the decision may be difficult for analysts to investigate and difficult for financial institutions to justify to affected customers. Explainable artificial intelligence seeks to make model predictions more understandable by identifying the factors that contribute to individual decisions or overall model behaviour. Ribeiro, Singh, and Guestrin (2016) introduced Local Interpretable Model-Agnostic Explanations (LIME), a method that approximates the behaviour of a model locally to explain individual predictions. Lundberg and Lee (2017) introduced SHapley Additive exPlanations (SHAP), a framework for attributing model predictions to input features using principles derived from cooperative game theory. These methods offer ways to investigate why a transaction has been assigned a high fraud score, for example, by identifying unusual transaction amounts, transaction frequency, or patterns of account activity where such variables are available in the dataset. Psychoula et al. (2021) specifically examine explainable machine learning for fraud detection and discuss practical considerations associated with model explanations, including the selection of reference data and computational costs. Explainability is therefore important not simply as a technical feature but as a means of supporting human review, auditability, and more informed investigation of suspicious financial transactions.

The development of fraud detection systems for low-resource financial environments requires consideration of more than predictive accuracy. Financial institutions and payment service providers may differ in their access to labelled transaction data, specialised technical personnel, reliable computing infrastructure, and resources for maintaining complex analytical systems. A model that performs well in a large research dataset may be difficult to deploy where processing capacity is limited, data quality is inconsistent, or technical teams cannot readily maintain the system. These constraints make computational efficiency, ease of interpretation, data protection, and practical maintainability important design considerations. Sculley et al. (2015) explain that machine-learning systems can accumulate technical debt when their dependencies, data pipelines, and operational requirements are not properly managed. Although their work is not limited to financial fraud, it provides a useful basis for considering the practical challenges associated with deploying machine-learning prototypes. In the Nigerian context, the Central Bank of Nigeria’s Payments System Vision 2025 emphasises the importance of developing a secure and efficient payments ecosystem. A prototype designed for resource-constrained environments should therefore favour an appropriate balance between predictive performance, computational requirements, transparency, and implementation simplicity. This study will consider these factors when selecting candidate algorithms and evaluating the feasibility of explainable machine learning for mobile money fraud detection.

A major methodological challenge in fraud detection is class imbalance, which occurs when fraudulent transactions represent a small proportion of the total transaction records. Under such conditions, a model may achieve apparently high accuracy simply by classifying most transactions as legitimate while failing to identify a substantial proportion of actual fraud cases. Consequently, accuracy alone is insufficient for evaluating a fraud detection model. Researchers should consider precision, recall, F1-score, precision–recall curves, and other suitable metrics that reflect the costs of false positives and false negatives. He and Garcia (2009) review learning from imbalanced data and explain why conventional classification approaches may perform poorly when the distribution of classes is highly unequal. Dal Pozzolo et al. (2015) also discuss the challenges of adaptive fraud detection in changing environments, reinforcing the importance of realistic model evaluation. For mobile payment data, additional considerations include missing values, duplicate records, changes in transaction behaviour, and the risk of information leakage between training and testing datasets. Vijayanand and Smrithy (2025) investigate an explainable ensemble-learning approach for mobile money fraud detection using the PaySim simulated transaction dataset and SHAP-based explanations, illustrating how fraud detection performance and interpretability can be studied together. Their work provides a relevant methodological reference, although performance on simulated transactions does not establish equivalent performance on real Nigerian payment data. These considerations support the need for a prototype evaluation that reports class-sensitive metrics and clearly distinguishes the properties of the dataset used from claims about real-world deployment.

Nigeria’s digital payment ecosystem provides a relevant context for studying explainable machine learning because financial service providers must address fraud risks while maintaining the accessibility and reliability of digital transactions. A useful fraud detection prototype should identify suspicious transaction patterns and provide explanations that can help a human reviewer understand why a particular transaction has been flagged. Recent research on artificial intelligence and fraud prevention in Nigeria discusses the potential of machine learning for detecting suspicious patterns, improving analytical efficiency, and adapting to evolving forms of fraud, while also recognising infrastructure and governance challenges.

1.2 Statement of the Problem

The expansion of mobile payment services has increased the importance of identifying fraudulent transactions without unnecessarily disrupting legitimate financial activities. Conventional rule-based systems can identify known patterns, but they may fail to recognise new or changing forms of fraud that do not conform to predefined conditions. Bolton and Hand (2002) explain that statistical approaches can support fraud detection by identifying patterns in transaction data, while Phua et al. (2010) demonstrate the relevance of data-mining methods to a range of fraud detection problems. Despite these developments, automated fraud detection remains challenging because suspicious behaviour may be subtle, transaction patterns may change, and fraud cases may be relatively rare compared with legitimate transactions. In low-resource financial environments, these difficulties may be compounded by limited access to labelled data, computational constraints, and a shortage of specialised personnel. The problem addressed by this study is the need for a practical machine-learning prototype that can identify suspicious mobile-payment transactions while accounting for the operational constraints that may arise in such environments.

A second problem concerns the difficulty of interpreting fraud detection decisions produced by complex machine-learning models. Although a model may assign a high fraud probability to a transaction, a prediction alone may not provide sufficient information for an investigator to determine why the transaction was flagged or whether the result is reasonable. Limited explainability can complicate the review of suspicious transactions, reduce confidence in automated decisions, and make it harder to identify model weaknesses. Ribeiro et al. (2016) introduced LIME to provide local explanations for individual model predictions, while Lundberg and Lee (2017) developed SHAP as a framework for explaining model outputs through feature attribution. Psychoula et al. (2021) specifically examine explainable machine learning in fraud detection and discuss practical considerations involved in producing explanations. However, the availability of an explanation does not automatically make a model accurate or fair; explanations must be assessed for consistency, relevance, and usefulness to human reviewers. The problem is therefore to determine whether explainable machine-learning techniques can provide both adequate fraud detection performance and understandable transaction-level explanations in a prototype designed for resource-constrained conditions.

A further problem relates to the difficulty of evaluating fraud detection models when transaction data are imbalanced, incomplete, synthetic, or not representative of the intended operational environment. Fraudulent transactions may account for only a small proportion of available records, making accuracy an unreliable indicator of a model’s practical effectiveness. He and Garcia (2009) explain that class imbalance can undermine conventional learning approaches, while Dal Pozzolo et al. (2015) demonstrate the importance of accounting for evolving data patterns in fraud detection. Publicly available datasets can support reproducible experiments, but they may not represent the transaction characteristics, fraud mechanisms, customer behaviour, or operating conditions of Nigerian payment services. Vijayanand and Smrithy (2025) evaluate explainable ensemble learning on the PaySim simulated mobile money dataset, showing the relevance of combining predictive modelling with explanation methods. Nevertheless, results obtained from simulated data should not be treated as proof of effectiveness on real Nigerian transactions. This creates a need for an evaluation approach that explicitly addresses class imbalance, data provenance, leakage prevention, appropriate performance metrics, and the limitations associated with the selected dataset.

Although machine learning and explainable AI have been investigated for financial fraud detection, there remains a practical need to assess how these approaches can be combined in a lightweight prototype intended for low-resource financial environments and evaluated using data relevant to Nigerian mobile payments. Existing work provides methodological foundations for fraud detection, class-imbalance handling, and explanation generation, but performance results from other financial systems or simulated transaction datasets do not automatically establish suitability for the Nigerian context. Vijayanand and Smrithy (2025) demonstrate the integration of explainable AI with ensemble learning for mobile money fraud detection, while research on fraud prevention in Nigeria identifies the potential of AI and machine learning alongside implementation challenges.

1.3 Aim and Objectives of the Study

The aim of this study is to develop and evaluate an explainable machine-learning prototype for detecting mobile money fraud in low-resource financial environments using anonymised or publicly available Nigerian mobile-payment transaction data.

The specific objectives are to:

1. Develop a machine-learning prototype for identifying potentially fraudulent mobile-payment transactions using suitable transaction data.

2. Evaluate the predictive performance of the prototype using appropriate metrics, including precision, recall, F1-score, and precision–recall analysis.

3. Integrate an explainable artificial intelligence technique, such as SHAP or LIME, to identify the transaction features contributing to individual fraud predictions.

4. Assess the prototype’s computational efficiency, interpretability, and practical feasibility for fraud detection in low-resource financial environments.

1.4 Research Questions

The following research questions will guide the study:

1. How can a machine-learning prototype be developed to identify potentially fraudulent mobile-payment transactions using the selected dataset?

2. How effectively does the prototype detect fraudulent transactions when evaluated using precision, recall, F1-score, and precision–recall analysis?

3. How can explainable artificial intelligence techniques identify the transaction features contributing to the prototype’s fraud predictions?

4. To what extent is the prototype computationally efficient, interpretable, and feasible for use in low-resource financial environments?

1.5 Research Hypothesis

The following null hypothesis will be tested at the 0.05 level of significance.

H₀: There is no statistically significant difference in fraud detection performance between the selected explainable machine-learning model and the baseline model when evaluated on the same test dataset.

The baseline model and the explainable model will be evaluated using an appropriate performance measure, preferably the area under the precision–recall curve or another metric justified by the class distribution. A suitable paired statistical test will be selected based on the evaluation procedure and data characteristics.

Methodological clarification: If the explanation technique is added to an existing predictive model without changing its predictions, explainability alone will not produce a statistical difference in predictive performance. In that situation, the hypothesis should compare two predictive models, while the usefulness of the explanations should be assessed separately.

1.6 Significance of the Study

The study is expected to benefit financial service providers, fraud analysts, software developers, researchers, regulators, and users of digital payment services.

1.6.1 Financial Service Providers

The study may provide a practical framework for investigating how machine learning can support the identification of suspicious mobile-payment transactions. The prototype may help financial service providers understand the trade-offs between predictive performance, computational requirements, and interpretability. Its findings may also inform decisions about whether lightweight machine-learning approaches are suitable for further testing within their fraud monitoring processes.

1.6.2 Fraud Analysts and Investigators

The study may help fraud analysts understand how explanation techniques can provide additional information about why a transaction receives a high fraud score. By identifying influential transaction characteristics, explanations may support the prioritisation of suspicious cases and help investigators determine which transactions require further examination. The prototype will support investigation rather than replace human judgment or establish fraud conclusively from a model prediction alone.

1.6.3 Software Developers and Machine-Learning Practitioners

The study may provide a technical reference for developers interested in building explainable fraud detection systems using accessible machine-learning libraries and computationally efficient algorithms. It may demonstrate how preprocessing, model training, class-imbalance management, performance evaluation, and explanation generation can be combined within a reproducible prototype.

1.6.4 Financial Regulators and Policymakers

The findings may contribute to discussions about transparency, accountability, and responsible automation in digital financial services. The study may provide evidence about the benefits and limitations of explainable fraud detection, helping inform consideration of systems that support human review, maintain appropriate records, and protect sensitive transaction information.

1.6.5 Researchers and Academic Institutions

The study may contribute to research on explainable artificial intelligence, imbalanced classification, mobile-payment fraud detection, and machine-learning applications in resource-constrained environments. It may also provide a foundation for further studies using larger Nigerian datasets, alternative algorithms, or operational evaluation with authorised financial institutions.

1.6.6 Users of Mobile Payment Services

Indirectly, the study may contribute to efforts to develop more reliable fraud monitoring systems. Better detection and review processes may help reduce the likelihood that fraudulent transactions go unnoticed. At the same time, appropriate evaluation of false positives is important because overly aggressive fraud detection can delay legitimate payments or inconvenience customers.

1.7 Scope of the Study

The study will focus on the development and evaluation of an explainable machine-learning prototype for mobile money fraud detection in low-resource financial environments.

Technical scope: The prototype will use one or more suitable classification algorithms, potentially including logistic regression, decision trees, random forests, or gradient-boosting models. The final choice will depend on dataset characteristics, computational requirements, and comparative performance.

Data scope: The study will use anonymised Nigerian mobile-payment transaction data where authorised access is available. If suitable Nigerian transaction data cannot be obtained, an appropriate publicly available dataset will be selected. A dataset such as PaySim may be considered, but its simulated nature must be clearly disclosed. Results from simulated or non-Nigerian data will not be presented as proof of effectiveness on real Nigerian transactions.

Explainability scope: The prototype will incorporate SHAP, LIME, or another justified explanation method to examine the contribution of transaction features to model predictions. Both individual transaction explanations and overall feature-importance patterns may be evaluated.

Performance scope: Evaluation will include precision, recall, F1-score, and precision–recall analysis. Additional measures, such as confusion matrices, ROC-AUC, inference time, memory consumption, and model size, may be included where appropriate. Accuracy will not be used as the sole indicator of success.

Operational scope: The study will examine computational efficiency and practical feasibility under resource constraints. It will not claim production readiness or live fraud prevention unless the prototype is separately validated in an authorised operational environment.

Methodological scope: The study will use a quantitative experimental approach involving data preprocessing, model development, training, testing, comparison with a baseline, and explanation analysis. Data privacy, class imbalance, information leakage, and the limitations of the dataset will be considered throughout the evaluation.

1.8 Operational Definition of Terms

Explainable Machine Learning: Machine-learning approaches that provide information about the factors contributing to model predictions, making the system’s outputs easier to interpret and investigate.

Mobile Money Fraud: Fraudulent or unauthorised activities involving mobile-based financial transactions, including suspicious transfers, account misuse, and other forms of transaction-related deception represented in the selected dataset.

Fraud Detection Prototype: A working experimental software implementation designed to classify transactions according to estimated fraud risk and provide explanations for its predictions.

Low-Resource Financial Environment: A financial setting in which computational infrastructure, financial resources, labelled data, technical personnel, or system maintenance capacity may be limited.

Transaction Data: Structured records containing transaction-related features, such as transaction type, amount, timestamp, and other available variables, subject to the contents and permissions associated with the selected dataset.

Anonymised Data: Data processed so that individual customers cannot reasonably be identified from the information retained in the research dataset, taking account of the possibility of re-identification from combined attributes.

Publicly Available Dataset: A dataset made accessible for research or educational use under its stated licence and conditions. Public availability does not automatically mean that the dataset contains real Nigerian transactions or that it represents all Nigerian payment services.

Model Explainability: The extent to which the reasons for a model’s predictions can be understood through techniques such as SHAP or LIME.

Precision: The proportion of transactions classified as fraudulent by the model that are actually labelled as fraudulent in the evaluation data.

Recall: The proportion of labelled fraudulent transactions that the model correctly identifies.

F1-Score: The harmonic mean of precision and recall, used to summarise the balance between these two measures.

Class Imbalance: A condition in which one transaction class, typically legitimate transactions, substantially outnumbers the other class, typically fraudulent transactions.

Computational Efficiency: The computational resources and processing time required to train the model, generate predictions, and produce explanations.

Prototype Feasibility: The extent to which the developed system demonstrates acceptable predictive performance, interpretability, and computational practicality for further investigation in a resource-constrained setting.

Project – Explainable Machine Learning for Detecting Mobile Money Fraud in Low-Resource Financial Environments.