What is our primary use case?
We used it for an Academy project and we also used it for a credit card transaction fraud detection project. The purpose of the project was to denote anomalies or customer records that were different from usual.
The reason that we used this solution was that it was the only end-to-end framework available on the market at that time. If I were to build my own solution on AWS, I would have had to bring in all the components, but AWS Fraud Detector has the capability to blend in not just the data streaming part, but also the multi-building part, ultimately chiming in with Amazon QuickSight for the visualization and Amazon SNS for notification.
How has it helped my organization?
In machine learning, we may measure a metric, something called recall — the other metric is false-positive. A recall refers to the total amount of fraud — how much were we able to recall (capture). Overall, we got some really good results. We got roughly a 77% recall, which meant 77% of the total fraud was actually picked up by Amazon Fraud Detector. We could pick a lot of fraud at the same time, but you're false positive, so the number of customers who may not be fraudulent, who get tagged as fraudulent also increased. We were able to reduce the false positives to roughly 20%, which is an extremely small number compared to other solutions or other models in place.
What is most valuable?
In terms of infrastructure — seamless integration was the most valuable feature. Everything was really plug-and-play style. I could see what I had to do, and I could customize it according to my needs.
The second valuable feature involved the machine learning part of things where I could experiment with different algorithms in Amazon — Amazon SageMaker and experiment with different algorithms.
What needs improvement?
The problem I was facing, from a machine learning perspective, it only had a supervised learning capability. You would have to provide your data live, but in fraud, the pattern of the fraudsters keeps changing and it's impossible to provide data labels. That's where the user unsupervised learning comes in handy — you don't have to tell them, "okay, this is fraud and this is not fraud."
If unsupervised learning was also incorporated with Amazon SageMaker, that would be really cool. I am talking about anomaly detection algorithms, like isolation, forest, or anything on the neural network side for anomaly detection, including autoencoders. These are some things which companies would really like to use.
There was also a problem with latency. In fraud detection, everything needs to be happening in real-time, but some of the algorithms ran for three to four minutes, which is not a viable option.
For how long have I used the solution?
I used this solution for six months.
What do I think about the stability of the solution?
It is stable. I didn't face any of the issues with respect to crashing or it being nonresponsive, or the server being down.
What do I think about the scalability of the solution?
I would say it is quite scalable. The services are built on top of each other, so you could seamlessly bring in any of the services and get them into play — it is highly scalable.
How are customer service and technical support?
The technical support was pretty responsive. Some of the things required access to Amazon, instances like a higher compute capacity and all that, but they went on doors and they were responsible in real-time. For some of the things which required greater access, they took not more than one day.
How was the initial setup?
The whole process was straightforward, actually. Amazon has really good documentation. I have to give credit to them on that. It was plug-and-play, and we knew we had access to all the codes and the repositories on GitHub — it was seamless.
The entire setup just took a couple of days.
What other advice do I have?
Make your integrations right and test them properly because it is still a new product. Things could break depending on their use case. If you are able to ensure that everything is one hundred percent connected and know that nothing is going to break — that would be good.
On a scale from one to ten, I would give Amazon Fraud Detector a rating of eight.