We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
A semi-supervised Anti-Fraud model based on integrated XGBoost and BiGRU with self-attention network: an application to internet loan fraud detection.
- Authors
Gorle, Venkata Lakshmi Narayana; Panigrahi, Suvasini
- Abstract
Recently, fraud debt has been one of the major issues for Internet financial institutions. Due to fraudulent activities, huge losses are occurring in financial institutions. Hence, there is a need for a method of analyzing and detecting fraudulent transactions and separating them from genuine ones. Supervised learning approaches are mainly used for fraud detection since they consider the fraudulent set, which can be known from past transaction analysis. Though these models are interpretable, the prediction accuracy of these models remains challenging. However, these approaches fail to perform well when there are changes in customer behaviour. Moreover, it is complex to identify abnormal transactions due to data imbalance. Hence, this work presents a semi-supervised outlier score-based Anti-Fraud model to identify the loan applicant as a genuine or fraudulent debtor. The proposed work has the stages like a pre-processing module, Data augmentation and classification model. After per-processing the data, different outlier models such as Z-score and Isolation forest (IF) are applied to generate more data. Then, the Unsupervised K-Means Clustering (KMC) granularity-based Outlier scoring method is proposed to augment the datasets with too many scores. This clustering module clusters the loan applicants based on their credit history. Then, the Z-score and IF are applied to each cluster to augment the original dataset with different scores. This normalized data is input to the XGBoost-bidirectional Gated Recurrent unit (BiGRU) self-attention network (SAN). This XGB-BiGRU-SAN is used to capture more efficient dynamic information. Further, a mathematical model, an Arithmetic Optimization algorithm (AOA), is used to optimize the network weights. The performance of a proposed XGB-BiGRU-SAN Internet loan fraud detection is analyzed on the two benchmark datasets, like the leading club and bank loan status. The proposed XGB-BiGRU-SAN achieved better classification accuracy, precision and recall of 99.05%, 99.11% and 99.34% on the leading club dataset. Further, the accuracy, precision and recall values achieved in the bank loan status dataset are 98.67%, 98.82% and 98.62%, respectively.
- Subjects
INTERNET fraud; FRAUD investigation; OPTIMIZATION algorithms; BANK loans; K-means clustering; SUPERVISED learning
- Publication
Multimedia Tools & Applications, 2024, Vol 83, Issue 19, p56939
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-17681-z