Artificial Intelligence

Why are recommender systems so bad?

April 27, 2021

You would think it would be easy. A store-front application stores your purchases in a database. After some period of time, or some number of purchases, the artificial intelligence backend of the store-front application begins to make recommendations regarding your upcoming purchases. If your last six purchases for t-shirts have been for 1990s metal bands, say, then Pantera and Iron Maiden would appear in recommendations for logos more often than would Hanson or the Backstreet Boys.

So far, so good; if you go to a bookseller’s website or to a music shop, it is very likely that the AI will learn your preferences regarding books or music quickly, since you probably have only one or two categories which you enjoy, and the overlap doesn’t lead to confusion. A fondness for rockabilly music, plus occasional purchases of gospel music, won’t lead to much confusion. 

The reason for this is that music has clearly-established genre boundaries. In the AI world, we would call this an ‘ontology’. An ontology is a set of categories in a subject domain that shows the relationships between things. A music shop might sell records, CDs and clothing; within the records and CDs, might sell rock, classical and jazz, blues, folk, country and soul.

The problem arises for huge vendors like Amazon, where things can get muddled. In the orader world of the biggest vendors, who strive to sell everything, or at least to serve as a storefront for vendors who sell anything, there isn’t a general ontology in place for all world goods. Efforts are underway to create such an ontology, but that’s obviously a massive undertaking.

So, at least for a while, liking the band Deep Purple may mean that red onions are more prominent in your Whole Foods search than yellow onions. And, as they say, “that’s not a bug -- that’s a feature.”

Published by

Curious to know more?

Connect with our experts to have a look at your business challenge.

Book a demo

Download whitepaper

Learn about how we apply our science to real world assignments.

Download whitepaper

You might also be interested in

Get the answers to your strategic questions in just one day.*