Learning Imbalanced Datasetsusing DifferentiatedRegularization of the Loss Function

A secure and efficient banking system is a critical prerequisite for the economic growth of any country. In the era of e-banking, banking industry plays a vital role in the functioning of organized money markets, and also is medium mobilizing funds and channelizing them for productive purposes. It h

Project Title

Learning Imbalanced Datasetsusing DifferentiatedRegularization of the Loss Function

Project Area of Specialization

Artificial Intelligence

Project Summary

A secure and efficient banking system is a critical prerequisite for the economic growth of any country. In the era of e-banking, banking industry plays a vital role in the functioning of organized money markets, and also is medium mobilizing funds and channelizing them for productive purposes. It has been seen during the last several decades that even the solid markets and long functioning banking systems have had significant bank failures and bank crisis on account of increasing magnitude of frauds and scams. Every year billions of dollars are lost due to fraudulent transaction and according to recent reports it is expected that by 2020 only the global card fraud will exceed $35.54 billion. So there is a dire need of a system that can help us mitigate the problem of fraudulent transaction in banking industry.

We propose to build a fraud detection system based on deep neural networks that can detect and alert us for fraudulent transactions. The usual limitation of such system based on deep learning is deep neural networks doesn’t perform well on predicting fraudulent transaction which is due to the fact that only 5-6% of the all transaction are fraudulent. In deep learning jargon this scenario is known as class imbalance problem. To overcome this class imbalance problem and get satisfactory result we are proposing our variant of deep neural network that has a special loss integrated in it, that could stop the usual degradation in performance caused by class-imabalance (frauds are only a minority of all transactions). After integrating our loss with the traditional neural network we are able to build a classifier that performs significantly better on detecting fraudulent transaction. Hence, now a system can built over this classifier that be used to detect fraudulent transaction and save billions of dollars globally.

Project Objectives

The project objectives are two-fold. Firstly, we aim to build a fraud detection system which is able to detect fradulent transactions much more effectively than the current state of the art. Secondly, we aim to investigate the effectiveness of our proposed methodology on class imbalance in general so that  we can propose a novel technique using Deep Learning to deal with the problem of class imbalance based on cost-sensitive and discriminative feature learning.

Project Implementation Method

As mentioned above, we propose to build a fraud detection system based on deep neural networks that can detect and alert us for fraudulent transactions. Specifically, we aim to synthesize a novel loss function which will lead to an end-to-end solution to the problem of class imbalance. We aim to merge two of the major paradigms in Deep Learning, Cost-sensitvie learning and discriminative learning. We wil evalute our method on six highly imbalanced publicly available datasets of credit fraud detection. We will also evaluate our method on real datasets if they become available.

Benefits of the Project

Anomaly detection is a very common problem which has been seen in several domains such as Banking, Insurance,  Cloud security, medical data. Our project can help in improving the current performance of  machine learning systems in determining these anomalies.  Moreover, our project can work as an end-to-end framework which is not seen commonly in these applications. 

The expected outcomes of the project are as follows:

1. a software that analyses the tranactions database  at a bank for fraud cases (one out of a dozen in this POC, the Focal HiRisk Customers).

2. Benefit to banks which lose face and money. Even a 5% saving in the of the money being lost due to fraudulent transaction has a significant global impact. Audit firms have to be paid by the bank for threshold re-calibration, which may be another cost saving.

Technical Details of Final Deliverable

Our final deliverable is an end-to-end framework for credit fraud detection which will analyze customer transactions and detect frauds. This system can easily be deployed on any Bank system. One of the major merits of this project is that our framework can be used for numeric as well as image based data and is very easy to integrate irrespective of the format of the data.We will provide detailed results of our framework on the datasets that we have been using for evaluation. We will also provide comparision of our framework with the models that are being used currently in the domain of fraud detection.

Our model will be implemented on the Pytorch framework which is currently the simplest and most commonly used Machine learning framework. The project can further be extended to provide cloud based service using Amazon Web services(AWS).

Final Deliverable of the Project

Software System

Type of Industry

Finance

Technologies

Artificial Intelligence(AI)

Sustainable Development Goals

Industry, Innovation and Infrastructure

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com