How Amex Deals With Fraud Detection Using RNNs?
This blog post summarizes what I learned in a presentation at NVIDIA’s GPU Technology Conference (GTC) by Manish Gupta, Vice President of Machine Learning & Data Science Research at American Express and Abhishek Khanna, Vice President of Fraud Risk Decision Science at American Express titled “Fraud Prevention in Real Time Using Recurrent Neural Networks”, as well as my own my own studies and experience.
The American Express Company (Amex) is an American multinational financial services corporation. American Express has 64000 worldwide employees and $1.24 trillion business is made by them every year. American data has a global presence across the entire chain of finances and 100% of the models are ML and big-data powered.
Why Big-data?
Millions of buyers and millions of sellers together making billions of transactions lead to over a trillion combinations and dollars in transactions a year in charge volume. Hence they must try to provide customers with a safe way to pay, to reduce the amount of fraud, and create trust within the customers.
How does Machine learning Impact Business at AMEX?
- Customer Management & Loyalty marketing: Retaining existing customers and growing by the implementation of strategic and useful incentives. This includes introducing the latest AI/ML technologies to keep their customers happy (For example, introducing Cognitive process automation, Robotic automation of processes, or personalized recommendation using predictive analytics).
- Credit Line assignment: A credit line is the amount of money that can be charged to a credit card account. The size of a credit line, and how much of it has been borrowed, have a large influence on consumer credit scores. Machine learning and AI models capture non-linear relationships and give high predictive accuracy. They also leverage far more number of features without sacrificing robustness.
- New-Customer Marketing: Attracting new business by building your trust and building your brand.
- Customer Service: Improving Customer service by building a more robust automated customer support to reduce waiting time for each customer.
Did you know that only 9% of consumers believe it’s acceptable to wait up to one minute to speak with an agent?
Fraud Detection:
Fraud is a complex problem. Various types of fraud include Online, counterfeit, Account takeover, lost/stolen, Non-received, and fraud applications. To increase security and reduce risks of fraud, the obvious solution seems to be adding multiple layers of security. But, the increased passwords may seem to be cumbersome and may hamper customer experience. Each of these frauds must be balanced out with customer experience. Hence the point is to take advantage latest deep learning techniques and hence improve the risk models.
Amex deals with fraud detection at the time of the transaction was made. The data has millions and billions of data hence proper infrastructure is required to handle this data.
AmexData: Over 1 Trillion transactions every week.
ImagenetData: Over 14 million Images ( Most-Used dataset for Computer vision).
As listed above, We must understand the scalability of data present and we need to have RNN come up with better architecture when it came to solving the problems. Fine-tuning is needed. The pick between transfer models or segmented modelling for online vs/offline models as to which model should be used for each case as well as fine-tune these architectures and hyperparameters of the data.
Deep learning is very popular for homogenous data and unstructured data ( dealing with arrays in case of images, or voice images ) but Amex has heterogeneous data and structured data. So the bank-balance might dwell in a large range based on different customers. It also has lots of categorical data like return payment or delinquency. All of this results in huge dimensionality. In Amex, there are about 100 of features which take up lots of values, and hence while creating a customer’s behaviour, the dimensionality blows up.
Let us examine the behaviour of a consumer if he has a credit card. He goes to the restaurant, then to the grocery store, later to the theatre, hence returns home with a taxi, goes to have dinner outside, fills petrol in his own car and goes to an amusement park with his family on weekends. Hence the number of categories and number of places and how much is spent, this information for each customer over a particular period of time can increase dimensionality a lot.
All this information is taken together and then LSTM based RNN is used to provide a fraud probability score. The models compare sequences and behavioural patterns of and if the sequence is diff when it comes to the particular person, then the alarm is increased.
If the frequency of transaction increases k-fold, where k is the threshold value at which the model predicts fraudulent activity or the transaction is in a different country or with a different IP address, then the fraud rate increases. When risky payments are done at late-night, especially online, an alarm is raised.
A normal NLP problem deals with 30k-40k text variables while the RNN model here deals with millions of such factors, increasing continuously over time.
Now, How does one control the dimensionality of the model? Tokenisation comes into mind, and there are multiple tokens required to be prepared. Dimensionality must have discrimination. For RNN grand variety is a challenge, i.e how many sequences must be taken and how often they must be taken, whether a month, a week, before applying it on the model. Hence the prior time is also one of the factors that must be taken into consideration.
Intuition: One thing while solving Fraud detection is the need to understand how exactly this deep learning model works and bring intuition on a human level. To bring about this intuition, various models are simulated to bring about how the RNN will work and react to different scenarios. This simulation might be used along with adversarial attacks. In adversarial attacks, inputs designed by applying small intentional perturbations to original training examples or inputs leading to an incorrect classifier prediction with high confidence.
There is a huge need for real-time deployment with low latency. Hence Nvidia partnered with AMEX to build RNN real-time solution for TensorRT. Go on the below link for more information.
American Express Adopts NVIDIA AI to Help Prevent Fraud and Foil Cybercrime
The advantages of the RNN based real-time deployment With Nvidia is as follows:
- Faster integration of Deep learning Algorithms in the market.
- Scalable-multiple Machine Learning models can be implemented at the same time.
- GPU acceleration: Helps to reduce latency and give real-time feedback on fraud detection in a matter of seconds.
- Fault tolerance and redundancy with auto-recovery of the models and systems implemented.
- Maintainable: Modular design and contemporary benefits.
Check out the NVIDIA GTC which has brought me to the infinite wandering of knowledge on Deep learning, GPUs, AI, and Machine Learning.
Please do add me on Linkedin and follow me on Medium for more DL, AI blogs!
