Fraud and Financial Crime Update (AI and Machine Learning in Banking – Part 1)
Attention fraud fighters! For the latest fraud and financial crime updates around AI and machine learning in banking, responsible AI, and fraud risk management, please subscribe: https://www.youtube.com/@Feedzai.RiskOps
CHAPTERS:
0:00 AI and Machine Learning in Fraud Prevention and Detection
0:44 Supervised Machine Learning: Tree-Based Models and Random Forests
1:56 Boosting Models
2:21 Probabilities and Reason Codes
TRANSCRIPT:
With the recent news about the UK actually pushing out a report about the five core principles of AI ethics, today we're going to talk about the AI machine learning application on fraud prevention.
My name is Xin and this is your Feedzai Financial Crime News Weekly Update.
In machine learning, specifically for fraud prevention and fraud detection, the most used technique is actually supervised machine learning model. The reason being that a lot of times you already have a fraud operation team, as well as some sort of a compliance, as well as feedback loops that are coming in, so you would already have a decent understanding of what kind of fraud you are experiencing. And what you really need is actually using the machine learning algorithm to be able to help you to better detect fraud.
The supervised machine learning model, what it does is actually make sense with thousands of combinations of different characteristics and then tell you, collectively, what the suspicious activity looks like. And the most commonly used algorithms are usually tree-based models.
There are a couple of algorithms that are very famous - random forest is one. You can think about it as each of the trees is making a decision. It’s similar to a traditional decision tree, right? So how the decision tree works is basically I look at the timing of the transaction. If it's at a certain time, I think it's more suspicious, so it goes to one bucket.
And then after that, I do a further split saying, “Maybe look at this transaction location, I also think it’s more risky.” And then I'll look at, oh, this email address also looks more risky. So you can actually split the tree. That's how we say it, you split the tree into different leaves.
At the end of the tree, you will be able to actually come to a decision saying whether this transaction is fraudulent or not.
And the random forest is actually a collection of trees. So essentially for each tree, you make a decision, and then at the end it's like voting, right? So the whole forest voted on whether you think this transaction is fraudulent or not.
Another very well-used technique is a boosting model - so things like XGBoost and LightGBM.
Every time you learn from last time what you have left - the residual - and then make cleverer and cleverer decisions at the end.
So when you actually use machine learning algorithms, the ones that I talked about - the tree algorithm - to make a decision on top of your transaction, what you get instead of a black and white, yes or no, which is really triggered by the rules, you actually can get a probability.
But not only are we giving you a score, we're also giving you reason codes. We can also call it white box explanation. So that reason code will also tell you what are the top contributed attributes to make the model think this transaction is more suspicious than the others.
Next time we'll talk about the model fairness in fraud prevention. This is Xin, and thank you so much for watching the Feedzai Weekly Update.
Attention fraud fighters! For the latest fraud and financial crime updates around AI and machine learning in banking, responsible AI, and fraud risk management, please subscribe: https://www.youtube.com/@Feedzai.RiskOps
CHAPTERS:
0:00 AI and Machine Learning in Fraud Prevention and Detection
0:44 Supervised Machine Learning: Tree-Based Models and Random Forests
1:56 Boosting Models
2:21 Probabilities and Reason Codes
TRANSCRIPT:
With the recent news about the UK actually pushing out a report about the five core principles of AI ethics, today we’re going to talk about the AI machine learning application on fraud prevention.
My name is Xin and this is your Feedzai Financial Crime News Weekly Update.
In machine learning, specifically for fraud prevention and fraud detection, the most used technique is actually supervised machine learning model. The reason being that a lot of times you already have a fraud operation team, as well as some sort of a compliance, as well as feedback loops that are coming in, so you would already have a decent understanding of what kind of fraud you are experiencing. And what you really need is actually using the machine learning algorithm to be able to help you to better detect fraud.
The supervised machine learning model, what it does is actually make sense with thousands of combinations of different characteristics and then tell you, collectively, what the suspicious activity looks like. And the most commonly used algorithms are usually tree-based models.
There are a couple of algorithms that are very famous – random forest is one. You can think about it as each of the trees is making a decision. It’s similar to a traditional decision tree, right? So how the decision tree works is basically I look at the timing of the transaction. If it’s at a certain time, I think it’s more suspicious, so it goes to one bucket.
And then after that, I do a further split saying, “Maybe look at this transaction location, I also think it’s more risky.” And then I’ll look at, oh, this email address also looks more risky. So you can actually split the tree. That’s how we say it, you split the tree into different leaves.
At the end of the tree, you will be able to actually come to a decision saying whether this transaction is fraudulent or not.
And the random forest is actually a collection of trees. So essentially for each tree, you make a decision, and then at the end it’s like voting, right? So the whole forest voted on whether you think this transaction is fraudulent or not.
Another very well-used technique is a boosting model – so things like XGBoost and LightGBM.
Every time you learn from last time what you have left – the residual – and then make cleverer and cleverer decisions at the end.
So when you actually use machine learning algorithms, the ones that I talked about – the tree algorithm – to make a decision on top of your transaction, what you get instead of a black and white, yes or no, which is really triggered by the rules, you actually can get a probability.
But not only are we giving you a score, we’re also giving you reason codes. We can also call it white box explanation. So that reason code will also tell you what are the top contributed attributes to make the model think this transaction is more suspicious than the others.
Next time we’ll talk about the model fairness in fraud prevention. This is Xin, and thank you so much for watching the Feedzai Weekly Update.
