Will recommendation systems always suck?
OK, there’s nothing like a leading question to get things going in the morning.
I have made several forays into the world of recommendation systems and machine learning both as a user and as a developer. As a user I have been utterly underwhelmed, every time.
There are several possible outcomes for a recommendation system:
- Recommends something you already have (obviously this is a degenerate case and easily filtered)
- Recommends something you’ve previously tried and rejected
- Recommends something you’ve never tried but don’t like
- Recommends something you’ve never tried and end up liking
Here’s an example. I like a band called Great Big Sea. The mere act of including this band into my declared content almost automatically generates a pile of recommendations that fall firmly into buckets 2 and 3. This is probably because there are a lot of people so desperate for a Great Big Sea sound that they’ll wander into a world of Irish folk music that has about as much in common with Great Big Sea’s essence as the band Phish has with marine lifeforms.
Intuitively I believe that recommendation systems will converge towards clusters. Even if you occupy more than one cluster you’re going to get the middle of the road for that cluster. Some people like that cluster. They may also like the manufactured pop band of the moment. While I don’t begrudge them that choice, they hardly need a recommendation system for that.
Today I read an article that suggests that recommendation systems can increase diversity for individuals but they do so at the cost of overall diversity. It suggests that diversity comes from increasing the friction of recommendations not decreasing them.
I once built a playlist and music sharing system based on the idea that you start out knowing what song you want to hear and then enjoy the serendipitous discovery of the songs that come next on someone else’s playlist. It’s like channel surfing on the radio. You stop when you recognize a song that you like, not one you don’t recognize. Then you stay on that station until they play too many ads or play a song you really don’t like. If you’re lucky you’ll hear a song that you didn’t know but really like. Unfortunately our system never saw the light of day but at least I found a few new songs.