Practical Application To My Watch Shopping Story

Earlier I posted What have you bought for me lately story here on my blog. My good friend, Tom Breur, Principal of XLNT Consulting in the Netherlands was quick to point out the following fascinating facts. Here is what Tom emailed me and I am publishing it here with his permission. Thanks Tom.

That story actually serves multiple purposes, depending on which point I
am making.

World wide, about 100 Million transactions per day, 50 Billion transaction per year are being screeened online
for fraud, using a Neural Network model (in combination with business
rules). The market leader is Falcon, owned by Fair Isaac, formerly HNC.

Learning points:

1 – Because of the huge volume, the learning is remarkably precise.
Conversely, you need sufficient volume to tease out the signal from the
noise.

2 – You need the “feedback recording” of events in order to make such data
capture a sustainable organisatiopnal activity: credit card companies
“know” when they hit a false positive (they call, and then pass the transaction),
the false negatives are claimed (disputed) after the client receives their
statements. These data are fed back in to the application ot make it
“adaptive”

3 – There are costs for false positives (spoiled surpises, like yours 🙂
– that’s what everybody expects if you tell the story right which makes it
so much better), and false negatives, fraudylent transations that were
accepted. If you put those in the equation, you get “better” models, that
is, better tuned to the business needs. (another example of asymetric mis
classification costs I always use is credit cards: when accepted +$200
profit, when defaulting -$5000 write-off, therefore you cannot afford to
err very often).