Anomaly Detection Model for Imbalanced Datasets

TL;DR

This paper introduces a dynamic unsupervised anomaly detection method for financial frauds in imbalanced datasets, combining stochastic intensity models with Kalman filters to improve prediction accuracy.

Contribution

It presents a novel approach integrating stochastic intensity models and Kalman filters for fraud detection in highly imbalanced financial datasets.

Findings

Better predictive power in imbalanced datasets compared to existing models

Effective estimation of dynamic intensities using Kalman filters

Applicable to real-world financial transaction data

Abstract

This paper proposes a method to detect bank frauds using a mixed approach combining a stochastic intensity model with the probability of fraud observed on transactions. It is a dynamic unsupervised approach which is able to predict financial frauds. The fraud prediction probability on the financial transaction is derived as a function of the dynamic intensities. In this context, the Kalman filter method is proposed to estimate the dynamic intensities. The application of our methodology to financial datasets shows a better predictive power in higher imbalanced data compared to other intensity-based models.