Explainable Machine Learning for Detecting Fraudulent Online Job Postings in Ghana: A SHAP-Based Approach

Technical Summary: Explainable Machine Learning for Detecting Fraudulent Online Job Postings in Ghana

Problem Statement
Employment fraud, particularly through online job postings, is a growing concern in Ghana, characterized by deceptive practices that exploit job seekers through fake advertisements, upfront fee requests, and data harvesting. Traditional detection methods, such as manual moderation and periodic audits, are reactive, labor-intensive, and struggle to scale against the volume of daily postings. Furthermore, while Artificial Intelligence (AI) offers powerful detection capabilities, the "black box" nature of complex models hinders their adoption in sensitive domains like employment. Stakeholders require not only high accuracy but also transparent, interpretable reasons for flagging a posting as fraudulent to ensure trust, accountability, and regulatory compliance.

Methodology
This study proposes a framework combining supervised machine learning classification with Explainable Artificial Intelligence (XAI) to detect fraudulent job advertisements.

• Dataset: The research utilized a dataset of 499 job postings extracted from Jobweb Ghana, comprising 360 legitimate and 139 fraudulent instances. The data included features such as company name, job description, salary disclosure, application deadlines, employment type, and educational requirements.
• Preprocessing: The dataset was split into an 80% training set (399 rows) and a 20% test set (100 rows). Categorical variables were transformed using one-hot encoding to prepare the data for modeling.
• Models Evaluated: Five classification algorithms were developed and compared:
  1. Multinomial Naive Bayes
  2. Logistic Regression
  3. Random Forest
  4. K-Nearest Neighbors (KNN)
  5. Decision Tree
• Evaluation Metrics: Performance was assessed using Accuracy, Precision, Recall, F1-score, and ROC-AUC.
• Explainability: To address the interpretability gap, SHapley Additive exPlanations (SHAP) were applied to the best-performing model to quantify the contribution of each feature to the predictions.

Key Results

• Model Performance: All five models demonstrated strong predictive capabilities, with accuracy scores ranging from 0.88 to 0.92.
  • Multinomial Naive Bayes emerged as the top performer, achieving the highest Accuracy (0.92), ROC-AUC (0.9635), and F1-score (0.8621). It exhibited a strong balance between precision (0.8333) and recall (0.8929).
  • Logistic Regression also performed well with an accuracy of 0.91 and a robust ROC-AUC of 0.9457.
  • Random Forest achieved 0.92 accuracy but slightly lower ROC-AUC (0.9303) compared to Multinomial Naive Bayes.
  • Decision Tree showed the lowest performance (Accuracy: 0.88), suggesting potential overfitting or limited generalization compared to ensemble and probabilistic methods.
• SHAP Analysis Insights: The application of SHAP to the Multinomial Naive Bayes model revealed the following key drivers of fraud detection:
  • Educational Requirements: Features related to degree requirements (specifically "Low" and "Standard" requirements) were the most influential predictors, contributing significantly to the classification of a posting as fraudulent.
  • Transparency Indicators: The disclosure (or lack thereof) of application deadlines and salary information were identified as critical factors. The mere presence or absence of these details significantly influenced the model's output.
  • Job Categorization: The number of categories assigned to a job posting was a moderate predictor.
  • Limited Impact of Branding: Features related to company branding, such as the presence of a logo, company name, or contact details, had minimal impact on the predictions, indicating that content-based characteristics are more indicative of fraud than organizational presentation.

Key Contributions

1. Contextual Application: The study applies machine learning specifically to the Ghanaian context of online recruitment fraud, addressing a gap in literature regarding local employment scams.
2. Integration of XAI: Unlike many studies that focus solely on predictive accuracy, this work integrates SHAP to provide transparent explanations for model decisions, addressing the "black box" problem in fraud detection.
3. Feature Identification: The research empirically identifies specific, interpretable indicators of fraud (educational requirements, deadline transparency) that can be used for rule-based filtering or human review.
4. Model Comparison: It provides a comparative analysis of probabilistic, ensemble, and tree-based models, demonstrating that probabilistic approaches (Multinomial Naive Bayes) may be particularly effective for this specific categorical dataset.

Significance and Claims
The paper claims that explainable machine learning models can effectively detect employment fraud while providing the transparent insights necessary for stakeholder trust. The study demonstrates that:

• Machine learning techniques are viable and effective for distinguishing between genuine and fraudulent job postings in Ghana.
• Probabilistic models like Multinomial Naive Bayes offer an optimal balance of sensitivity and reliability for this task.
• Explainability is not merely an add-on but a prerequisite for practical deployment; SHAP analysis successfully translated complex model outputs into actionable insights regarding educational requirements and transparency signals.
• Fraudulent postings are more effectively identified through content-based anomalies (e.g., vague requirements, missing deadlines) rather than superficial organizational features.

The authors conclude that integrating these explainable models into recruitment platforms can improve decision-making for administrators and regulators, offering early warning mechanisms against online employment scams.