Hacker News Reader

Comment by maxerickson | Hacker News Reader

A binary score is still a score.

There's absolutely a difference between a binary "fraud/not fraud" flag and a continuous variable for quality of customer. They measure different things and have very different use cases.

Is there a perceptible difference to a consumer who has been flagged as "fraud" when they are not, in fact, someone who has committed fraud?

If not, then there's no difference.

They are literally two very different types of data. This isn't an ambiguous issue: One is a continuous variable, the other categorical. Whether or not one or the other could erroneously flag someone doesn't change that, and doesn't mean they're the same. Because the nature of those differences means the use cases are different. A continuous score has much wider versatility than a black & white binary value, and allows for more nuanced use. Even if it's just in the realm of fraud detection, a continuous score allows for more safeguards, e.g., if it's low but not too low then it triggers further review rather than outright rejection.

You were replying to a comment that said "a binary score is still a score"

The comment you appear to be arguing with, is completely true. A binary score is of course a score, and in fact that's almost always how credit and trust scores are actually perceived.

From the company's perspective many scores are continuous, but from the consumers perspective that's mostly a distinction without a difference.

Usually from the consumer side you're being told you got the job or didn't, or got the credit card, or loan, or apartment, or didn't.

I never said it wasn't a score, my issue was the implications that, both being scores, they were somehow equivalent. They are not. A continuous score may be used with a threshold to perform the binary categorization. In fact that is all but guaranteed. But the continuous score, as I stated, has more possibilities for nuanced use. The binary score is derived from something continuous.

And you aren't usually told you got the loan or not. That is one possibility, but the more likely one is the continuous score of the credit rating translates into a continuous score for the interest rate. This is why the distinction between a binary and continuous score is important: t This continuous assignment of interest rate isn't possible with a single binary variable.

Categorization into groups that you will or will not do business with is scoring. Pedantically explaining what sort of scoring it is doesn't change that.

Except that sort of will/won't is the only possible result of binary variables, while continuous allow more possibilities, like assigning a continuous variable interest rate on the basis of a continuous input.

I wasn't explaining some basic aspect of statistical variables for it's own sake, in this context it is not some pedantic hair splitting, it's a relevant aspect of the types of things that can be done with this data.

First off no one is saying "you look like fraud", they're saying details of your purchase looked suspicious. This is the first place where people need to pump the brakes, there is not some dire sinister plot, you went to make a purchase, it looked weird, we want to make sure it all checks out. This happens _all the time_ with credit card purchases, historically from the card issuer, but card issuers are actually encouraging retailers to be more vigilant in the face of massive increases in card fraud, so now it's distributed. Honestly, whenever I've had the minor inconvenience / world-concept-shattering experience of having a card transaction declined & needing to go through around 15 mins of rectification, I've actually appreciated that they are running systems to prevent rampant fraud, because it's an even bigger pain in the ass to clean up after that than it is to endure some extra information checks on rare occasion.

Yes, this is why I wanted to distinguish between a binary score and a continuous one. A continuous one allows for a flag for review at one threshold or an outright refusal at another. It's exactly what credit cards do. A purchase may go through, but it is suspicious so they call you to verify. Or it could be really suspicious so they block it and you have to contact them. It's happened to me both ways.

Seriously. If companies are relying on the fraud estimates, it isn't going to be fun to be a false positive.

I mean people have been subjected to fraud scoring for decades, you get declined, you follow up, and you get in w/ a new data point for the next time.

They've been subject to fraud scoring across a large group of disparate companies for decades?

I would say that's absolutely not true at all. And to the extent that it is true, via credit reporting agencies, it's extremely aggressively regulated to allow consumers to see and challenge how that data is used, as well as to completely opt out of that system.

Credit card companies use your past history across merchants in risk scoring transactions, yes, and they also pull in data from external sources that are outside of their walls to further enrich that data.

Anyways -- this is all massively orthogonal to what I originally wrote, which was around Sift being used as a consumer scoring system. I shouldn't have "taken the bait" so that's my bad.

Not when these new automated systems don't allow for a follow-up because it would be too expensive to code or hire a human to do it.

How many companies don't have a support@____ email address? Seriously, they'll figure it out, it's really not that large of a haul to get stuff corrected, and it's the difference between "we can't take credit card payments" and "a very low-percentage of real users have some extra friction / bounce somewhere else".

Right, it's not a score, it's just a thing you have to argue with them to change. Not a score though.

Reply on news.ycombinator.com