Yahoo Music Unlimited, Recommendation Engines and Serendipity

Photo: Michela Mongardi

Last night, I plunked down the plastic to get a Yahoo Music Unlimited To Go (would you like a non-fat soy latte with no foam and a double shot with that?) month-to-month subscription. I hadn't done it before, but there was a recent update to my mp3 player that made it compatible with the service.

Basically, for the price of a single CD every month, you have access to the million or so songs that they have in their library on up to 3 PC's and your portable device. For once, the terms are actually an exact match to how I listen to music. Any music I listen to that wasn't over the air radio is coming from either my laptop or my mp3 player.

After a quick perusal to ensure that it wasn't just 1 million remixes of Eminem and The Black Eyed Peas, I gave it a shot. While the software is disappointing, I'm beyond being surprised by audio software that is hard to use. If it wasn't for audio sync packages and scanner software, I wouldn't have a bottom to my usability calibration scale.

Digressions aside, I managed to pull down a few albums and get them onto the mp3 player, thus accomplishing the overall goal. This morning, I've been exploring the other bits of the Yahoo setup, which include a recommendation engine and thus the point of this post.

Recommendation engines are the most obvious use of recorded attention data. When you talk to some folks, you'd think that they're the only use, but it's clear that, at the very least, this space is how we're going to explore the tools for applying recorded data. Amazon really has pushed this along, as have services like AmigoFish, Last.FM and, to a lesser degree, Netflix have all been doing this kind of thing, though predominantly by asking you to rate items more than watching what you do.

That's understandable as the mechanisms aren't really mature for watching what DVD's you pull off the shelf and put into the player. At the moment it's easier (for the provider of the recommendations) to just ask people to do the tracking manually. Eventually, that will have to go away. When it does, the whole thing will explode in usefullness. Until then, they're still interesting.

At any rate, what I wanted to talk about with recommendation engines is the idea of serendipity as a variable in the equation. Oftentimes, when I talk about having the computer watch your actions and figure you out, people have a visceral reaction. They are often assuming that the default implementation is for the computer to take complete action: delete files directly, buy music for you, open programs for you, etc. And, were those kinds of things part of an implementation, you'd have every right to fear how it could run unchecked.

However, really good implementations will act more like a personal chef, personal maid, personal assistant, etc. than a parent. When you're a little kid, you have someone cook for you, drive you around, clean the bathroom you used, and otherwise took care of these kinds of things. However, they did so on their terms. You didn't have much input into menu selection (I know I didn't get pizza and ice cream constantly) or whether they decided to run by Target on the way home from school. And, when most of us were old enough, we left that behind, because we don't want someone else running the show. As adults, we have a pretty strong reaction to someone trying to be our parents.

If you had a personal chef, after they'd cooked for you for a while, they'd probably just start suggesting meals that they know you like. Really good ones would introduce a bit of serendipity, which is a critical variable in recommendation engines. They'd, for instance, say, because you like pizza, you might like a calzone. Now, depending on your level of comfort and how much serendipity you like, they might only suggest moving from canadian bacon pizza to a canadian bacon calzone (small level of serendipity) or might suggest moving from canadian bacon pizza to a Thai peanut calzone (large level of serendipity).

Really good systems will let you adjust the level of serendipity (and REALLY good ones will figure it out on their own) you're comfortable with. Yahoo
's music engine lets you do this (to a point), by toggling a few options to favor positively rated items, etc. This is gives you the serendipity factor. If you tighten it down, you'll only get songs from your favorite artists and that's it. Only songs you've already told it you like.

This is similar to listening to a classic rock or oldies station. You know all of the songs they'll be playing and there won't be anything new (by definition). If you allow a bit of serendipity, you'll start getting a new song here and there, much like stations that play "the best of yesterday plus the hits". And, if you open it up completely to serendipity, it's like listening only to the stack of CD's your music nut of a friend lends you, full of artists you've never heard of.

Any recommendation engine needs to have a level of serendipity (or you'll just get the stuff you've already seen) and Yahoo's is doing some of the right things to let you control it. From what I see so far, it (along with Last.fm) will be a central part in me finding new music without spending much time doing so.

And, if you're looking to figure out if there might be a service you like better than Yahoo's, check out TechCrunch's comparison of the services, which just came across my desk this morning.