FusionGuard: Integrating LSTM-GRU with Temporal Anomaly Detection for Robust Supply Chain Fraud Identification

Christopher, Jeni Keziah Alfrida (2025) FusionGuard: Integrating LSTM-GRU with Temporal Anomaly Detection for Robust Supply Chain Fraud Identification. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (984kB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

The study provides a solution to the multidimensional challenges and introduces FusionGuard, a hybrid fraud-detection model that implements sequential learning (also referred to as sequential modelling), anomaly detection and class-imbalance remedies in a unified framework. FusionGuard integrates both LSTM and GRU layers to learn long-term time-dependent and short-term dynamic variations in transactional data, and techniques such as SMOTE are used to deal with extreme class imbalance, while focal loss reduces prioritisation bias toward majority-class samples. The proposed model is primarily supervised, with conceptual support from anomaly-detection principles, thereby increasing its flexibility, and it can be applied to a variety of supply-chain datasets that contain mixtures of both labelled and unlabelled data. the FusionGuard methodology involves more than two layers of preprocessing and feature testing. The interquartile range method is used to remove outliers, correlation is analysed to determine high-impact features, and scaling is done using StandardScaler to maximise the performance of gradient-based learning algorithms. Exploratory data analysis shows the severity of the class imbalance, and it is used to decide on the modelling techniques (strategies) that can work best with class-imbalanced distributions. The results of its performance define a standard with which the sequential deep learning models can be compared. The experimental outcomes explain that the hybrid model can considerably lower the false negatives, one of the most severe measures in fraud detection, and show almost flawless accuracy and recall in detecting fraudulent transactions.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Niculescu, Hamilton UNSPECIFIED
Uncontrolled Keywords:	LSTM (Long Short-Term Memory); GRU (Gated Recurrent Unit); SMOTE (Synthetic Minority Over-sampling Technique); Undersampling; Credit Card; Fraud; Detection; Bias; Imbalance; Comparison; Logistic Regression
Subjects:	H Social Sciences > HG Finance H Social Sciences > HG Finance > Credit. Debt. Loans. Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning H Social Sciences > HD Industries. Land use. Labor > Business Logistics > Supply Chain Management
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	07 Sep 2026 10:12
Last Modified:	07 Sep 2026 10:12
URI:	https://norma.ncirl.ie/id/eprint/9853

Actions (login required)

View Item