First, thanks for reading the article. I’d cut it down from its initial 5k words but it’s still a pretty substantial read :)
Now, I’m not entirely sure what you’re referring to by “how much did it cost.” Do you mean how much does fraud cost Lyft, or do you mean how much does running the machine learning models and larger system cost? If it’s the former, I really shouldn’t and can’t reveal variable costs that Lyft isn’t making public. What I can say is that we are performing pretty well within the industry as far as I’m aware.
It’s hard to attribute fraud losses to each and every improvement in the long term. We run experiments and track fraud and growth metrics, of course. But due to the adversarial nature of fraud, a win today may not hold in half a year when fraudsters get around the new measures. In the two years, our team has brought down fraud rates to less than half of what it was. Obviously, the figures varies as our product reach increases and attract more attention from malicious actors old and new.
I may be quick to judge, but if you’re really asking for a quick two-liner like, “doing X in ML shaves off Y% in fraud rate” to summarize the article…there’s really nothing like that. Our story has been about improving the infrastructure that we have to empower us to deal with fraud in better ways. Or maybe we just don’t understand fraud and machine learning well enough.
