NWAGOR, O. PASCHAL (2026) ANOMALYIQ: A THREE-PHASE COMBINED ANOMALY DETECTION MODEL FOR BROKERAGE TRADING OPERATIONS EMPLOYING AUTOENCODER, ISOLATION FOREST, AND LIGHTGBM WITH SMOTE. Other thesis, Godfrey Okoye University, Enugu.
|
Text
AnomalyIQ_Correction.pdf Download (1MB) | Preview |
Abstract
Anomaly detection in financial fraud within the context of stockbrokerage in Nigeria and mobile money transactions continues to defy rule-based approaches, which are unable to respond to new patterns of fraud, unable to scale up with increasing transactions, and do not offer any mechanism for prioritization and explanation. This paper introduces the semi-supervised AnomalyIQ, which represents a three-stage architecture for hybrid anomaly detection and fraud classification, thereby overcoming the challenges posed by rule-based techniques. Stage 1 involves training a pure-NumPy Autoencoder solely using normal transactions and then producing an anomaly indicator based on reconstruction error set at the 95th percentile from the training errors. The second stage involves applying an isolation forest algorithm that uses random recursive partitioning to derive the structural independence of outliers. The third stage involves training a LightGBM classifier using SMOTE-balanced data. The meta-features of base features and stages one and two will be used to enable supervised fraud classification. The thresholds that yield the highest Precision and Recall scores in one optimization process are chosen out of 1,000 candidate values, while every detected anomaly is labeled as having Low, Medium, or High risk based on SHAP TreeExplainer feature-level explanations. Evaluation of the proposed system was carried out using two public benchmarks: Kaggle Credit Card Fraud Detection dataset (total number of records: 284,807; fraud rate: 0.172%) and PaySim Synthetic African Mobile Money dataset (total number of records: 6,362,620; fraud rate: 0.13%). For the Kaggle Credit Card dataset, the system achieved Precision 100.00%, Recall 98.67%, F1-Score 99.33%, and AUC-ROC 99.90% for 56,961 test records without a single false positive result, versus 496 false positives for the standalone Autoencoder model. For the PaySim dataset, precision 98.03%, recall 98.12%, F1-score 98.07%, and AUC-ROC 99.99% were recorded for 180,000 stratified test records, representing an improvement of 58.3 percentage points in F1-score for the two-stage unsupervised pipeline (F1=0.41). These findings serve as empirical confirmation of the system design hypothesis. The whole architecture can be accessed as a fully functional web application available under https://anomalyiq-rho.vercel.app
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Engineering, Science and Mathematics > School of Electronics and Computer Science |
| Depositing User: | MICHAEL MADUBUKO |
| Date Deposited: | 22 Jul 2026 13:14 |
| Last Modified: | 22 Jul 2026 13:14 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/6024 |
Actions (login required)
 |
View Item |
