A Lost and Found Algorithm, the worst kind of algorithm.
AI, machine learning, and algorithms are all the rage, but it’s still quite a silly choice for Lost and Found.
A brief history of AI:
2012- Google made a massive breakthrough by training its neural network to identify a cat, with 70% accuracy.
2018- DeepMind’s Alphastar blew away the e-sports world by being pretty good.
2021- Two-minute Papers is a great youtube channel; everyone should check it out. What AI can do these days is wild.
Those are the good bits, but I may have missed some highlights.
Why an algorithm is a bad choice?
The problem (well, more aptly, the current solution) doesn’t lend itself to efficient algorithmic matching. The high variance and information disparity in Lost and Found create more noise than signal, making more work for the customer service team.
Variance- dentures, a prosthetic leg, a wallet, a phone, keys, and a wedding dress have and will continue to be turned into the Lost and Found. The wide variety of items result in a wide variety of process and information entry, creating a difficult problem for any matching approach.
Information disparity- two separate parties reporting information about what was lost and recording information about what is found doesn’t line up.
The two parties
- the person (lost and found worker, police officer, {{insert human name}}, etc.) recording descriptions of the found property
- the consumer reporting what is missing
Since they aren’t the same person and have different perspectives, reports result in varying descriptions, including shape, size, color, or even omitting information entirely.
These two factors make it incredibly difficult for a machine, well-intentioned or not (watch this video about paperclips maximizers, it’ll blow your mind), to connect relevant information in any helpful way consistently. You end up with chaotic pairings that force the staff to say no a lot.
Of course, the apparent option is to simply match names, serial numbers, or other particular details that aren’t impacted by a perception bias. Sure, but that’s not a lot of stuff, and you need to prompt that information from both parties, and you don’t need a Machine Learning model for that.
Algos aren’t the worst option: better than playing telephone
Algorithmic matching isn’t fantastic, but it’s far from the worst; I’m talking about playing telephone.
Telephone- it’s like that child’s game of whispering a message into someone’s ear, and the more people the message is passed through, the more garbled the message becomes. You know that game, right? (it’s a real game; I am pretty confident of that)
But this is the telephone game with Lost and Found, often with actual telephones.
Telephone for Lost and Found
Step 1- the consumer tells a lost and found worker what they lost.
Step 2- the worker then looks through all the lost and found stuff for the property based on the description.
Fin.
Swap in voicemail or email as the medium, and you have the most common way Lost and Found is managed.
A truly terrible experience. A ton of extra work, unnecessary back and forth, poor communication, and bad outcomes. It’s difficult to get worse than that.
The “Loser” is the best algorithm
What’s the alternative?
Well, the person missing property, i.e., the “Loser,” is the most qualified option to look for that property; they have the distinct advantage of already knowing what it looks like.
So why not let them?
Theft… sure sure. You wouldn’t want people to go “shopping” in the lost and found. We’re not talking about turning the Lost and Found into Amazon; just making it a lot easier for consumers to identify their missing stuff and submit information to prove ownership; we call it Self-Service.
The staff can quickly confirm or deny these claims and focus on the best bit, getting items back to owners.
Interested in learning more or want to point out why my opinion is fundamentally wrong? Drop a line below, and I look forward to vigorous debate or polite chat (you decide).
Originally published at https://www.liffhappens.com on January 29th, 2021
