Not long ago I found myself at a major financial institution talking about one of their fraud detection systems. Over the course of the conversation I stumbled onto the fact they have over 10,000 rules in place to detect fraud ... and oh so proud they were.
On the surface that might sound “powerful and amazing.” Nonetheless, that struck me funny it did. 10,000 rules … WOW! That must be brittle, expensive, and one giant liability I thought to myself. Such a detection system would catch exactly 10,000 things, nothing more, nothing less. Every new discovery would lead to new rules. Over time as the rule library further bloats it would get harder to manage and probably get slower and slower. And by the way, how many people actually understand all those rules and their interrelationship? Then as those people move on, how hard is it to get new people trained up on all those rules? Will they still be bragging about their extensive rule library when they have 20,000 rules?
Imagine telling your kid one day to quit throwing rocks at cars. Only to realize the next day you have to tell them to quit throwing rocks at SUV’s. Then the coming days, you realize you must also tell your kid not to throw rocks at trucks, fire engines, and ambulances. Ummm … 4,172 rules later you must come up with new rules like “don’t throw cans of Dr. Pepper at trolley cars.”
How about: “Don’t throw things at other people’s stuff.”
As parents quickly discover, teaching a principle like this is a much better course of action. While certainly not a perfect principle, at least it would roll up hundreds of explicit rules and catch countless conditions you never thought of. And yes, maybe this simple rule needs to be extended e.g., “unless they are bad people doing bad things and they need to be stopped.” That way if someone is coming at them on a skateboard with a knife they know it is okay to throw a chair at them.
Now back to the real world and a real example from my past. Circa 1993 we were building the first NORA (Non-Obvious Relationship Awareness) system for a casino. In this system the first relevance rule was basically: “Tell me when the bad guy is the good guy.” This one rule was created to detect and alert for such things as: the slot club loyalty card member is banned from gaming (on the Nevada Gaming Control Board’s Excluded Persons List) or the job applicant is a known gaming felon.
The second relevance rule was: “Tell me when the bad guy knows the good guy.”
With just these two rules, the system started kicking out all kinds of valuable, unanticipated insight including one of my favorites: An alert surveillance room operator noticed a dude cheating on a roulette table … making bets after the ball fell (called “past posting”). Dealers are supposed to watch for this. But somehow today this dealer kept missing this obvious scam. Casino security detains the cheater. The dealer says “I can’t believe this happened to me, I am so embarrassed, you surveillance folks are sure doing a good job, it won’t happen again.” During the arrest processing, the cheating player provided a different last name and address than used by the dealer. Fortunately, the cheater provided his real home phone number which happened to be the same number that the dealer had used on her original employment application.
The dealer pretending, up to this point, to not know the player rolled-over in an instant and confessed when NORA popped off a real-time alert: “The cheater is related to the dealer.”
Behind the scenes this was data finds data followed by relevance finds the user. Relevance, in this case, based on the principle; alert when the bag guy knows the good guy.
Had we deployed a traditional rules-based alert system, there was some chance the specific rule – if the employee’s job application phone number matches an arrest record – might have been missed. But because NORA was engineered around principles we caught this colluding roulette dealer. Notably, we would have also detected this had they been connected via an emergency contact phone number. Or maybe the player’s loyalty club card’s original address provided when they signed up (and since changed) was the same address used on the employee’s original job application (but not present on her current payroll record).
Data triage systems, especially those that must detect ever-changing crafty adversaries, should be principle-based where possible; otherwise, you won’t be one step behind. You will be at two or more steps behind!
Principle-based decisioning systems may surprise you … in a good way.
MISC NOTES
1. Maybe some classes of systems need a zillion rules, like the space shuttle program, for instance. But, that is out of my field so I don’t know.
2. The notion that “principles outperform rules” probably applies to most, if not all, of the decisioning processes. For example, I would prefer to see feature extraction, entity resolution, relevance detection, filtering, and insight publishing algorithms leverage principles over rules wherever possible.
3. Just to be fair, many systems will still have to have some very specific rules – like any transaction over $10,000 must be reported to FINCEN, it’s a law. This being not much different than telling your child they have to be home by 9pm on school nights, period.
4. And if you get to 10,000 principles, you might want to focus on more abstraction.
OTHER RELATED POSTS:
You Won’t Have to Ask -- Data Will Find Data and Relevance Will Find the User