CEO Blog Series: Chapter 5: Something Different

And now for a change of pace: an AI use that has nothing to do with DAM. But it does reinforce the key concepts we have been talking about: “descriptors”, a “space” of dots, and “nearest neighbors”.

What if you were a huge online grocery retailer, and one of your customers orders a bag of rice. You want to be as useful to them as possible (and you want to sell them as much as possible!) so you want to put on their screen other things they might want to buy. But do you suggest something from Mexican cuisine, or Indian, or Japanese?

Well, you “know” this customer: you know their shipping address, and you might have a record of their previous purchases. What if you generated descriptors for each of their transactions, by feeding an AI engine their customer number, their address, the date and what they bought, AND you did that for all their neighbors, and put that data as a set of dots in your (multi-dimensional but conceptually 3d) space?

“Pat”, your customer, orders their bag of rice. You encode that transaction, put its dot in the cube of space, and then you look for its “nearest neighbors”, first for anything “Pat” has bought. If every time Pat orders rice they have also ordered tacos and cheese, that gives you an idea of what might be useful to suggest to them they also buy.

What if Pat is a first-time buyer and you have no history for them at all? Well you might do much the same thing, encode this purchase, put its dot in our space, and then grab nearest neighbors (in this case literally: remember part of what we encoded was their address?). You can then see what cuisine using rice is common with their neighbors, and try that. If the neighbors purchased tacos and cheese and jalapenos, suggest Mexican cuisine. If curry, ingredients for butter chicken, coriander, go with Indian cuisine. Or if sashimi, seafood salad, or soy sauce, suggest Japanese cuisine. You may guess wrong, but the more data points you gather, the more likely it is that you guess right. And over time, you will be likely to be increasingly accurate, and presumably increasingly profitable and useful to your customers!

Of course, sometimes a new recipe will show up with a new use for rice, and your system will not know about it yet. But as more people order that recipe’s ingredients, a cluster of dots will grow and your system will learn the new cuisine.

By the way, note the system does not need to guess Pat’s nationality, and indeed it would be a mistake to do so: Pat could be Japanese and love Mexican food. The system works solely off facts, hard data it “knows” from its internal transactions.

This is called a “recommender system” and is used to suggest music playlists, consumer products, restaurants, the next video you might want to watch, and suggest partners on dating sites, among scores of other uses. And now you know how they work!