tastes great, less filling
no editor ok’d this column

recommendations

Before David Hyman founded MOG, he worked as the CEO for Gracenote, which supplies the online database for many digital audio players. He has, as he says, "studied recommendations for a long long time," and he jokingly confesses that "this is all i think about all day, everyday." While MOG is an Internet-based system which relies on software to keep things running, and while calculations are constantly made at MOG to help collate and classify information (MOG keeps track of your entire music collection and what you play from that collection, then points out which other MOGgers share similar taste preferences to your own), MOG differs from most other recommendation systems by also working a human element into the equation (and it should be noted that MOG is far more than just a music recommendation site, but given my recent obsession with that topic, I'm focusing on that aspect here). As an example, you can designate certain MOGgers as "trusted," and the MOG will create a recommendations page that shows you the top songs or albums of your trusted friends (and, since MOG knows what you have in your own collection, the list of songs/albums will not include anything you already own, so the recommendations will theoretically at least be new to you).

(Sidenote: I hope if David happens to read this, that he corrects any mistakes in the above description. As always when it comes to technology, I am a hardcore user but am relatively clueless about inner workings. And David is good about correcting things … when one website ran a story about the creation of MOG and said it was based in Kensington, David signed on to note that MOG is in Berkeley. And when the writer then said Kensington, Berkeley, same difference, David corrected him again. Us good Berkelyans know the difference. If the flatlands/hills class split is one of Berkeley's most obvious but least-spoken-of idiosyncrasies, then Kensington is the place that even the hill dwellers can say is full of snobs. And, I should add, we've spent all of our 30+ years in Berkeley living in the flats, and in fact, as far as I know, MOG itself is in our neighborhood.)

Back to the point … I've been particularly tangent-filled lately. David has a post on his MOG ("MOG" meaning both the system as a whole and our individual homes … I have a MOG on MOG) where he talks about how MOG recommendations work, and why he doesn't trust A.I. recommendation systems. Given his work with Gracenote and MOG, and the previously mentioned obsession (all he thinks about, everyday), he's well worth hearing on the topic, even though readers of this blog might guess that my own increasing belief in A.I. recommendations lead me to some different conclusions from David. He's pretty convincing, though, when talking about the "Magic Button" on MOG that helps narrow down your recommendations:

the magic button on the recommendations page is just us showing you what the moggers who are most like you are listening to. i don't believe recommendations get any better than this.

yet, in many instances, the trusted mogs filter works better. those you've explicitly determined are your trusted sources. invite your friends, put them in your top 10 trusted mogs, and now you've got a page with instant access to the people who turn you on to music in the physical world! – to me, that's the best.

having studied recommendations for a long long time (5 years working at gracenote), i do believe that music discovery through people works better than amazon type collaborative filtering algorithms, digital signal processing, or editorially driven similarities

And later, after I offered a feeble attempt at sticking up for A.I.:

a.i is too self-referential. if you're 12 and listening to backstreet boys, amazon will tell you to get britney spears. but in the real world taste is way more complex. it's about emulation. about who you want to be. in the real world, when you're 12 and listening to backstreet boys, your friends older brother who is COOL to you plays you the rolling stones and next thing you know, you are listening to the stones. this DOES NOT happen with collaborative filtering. no phish head listened to pavement until trey (their god) told them he liked pavement….

let technology serve as conduit to those you've come to trust. it's good at that.

If David is right about this, I still represent a marginal case (what a surprise, Steven self-marginalizes again). Here's the thing, for me at least: while I am not quite friendless, I don't trust anyone's taste preferences. That's part of what I'm trying to get at with my various posts on the subject: don't take it personally if we don't share the same tastes, because there's no reason we should. (And the corollary, don't trust my taste preferences, either.) What I trust are artificially intelligent algorithms that make an unemotional, educated guess at what I might like. I have 16 "trusted MOGs" … among them are Michael Goldberg (long time music writer, Sleater-Kinney fan, and the main man behind the late lamented Addicted to Noise), Steven Levy (tech writer and former fantasy baseball mate), Michael Snyder (longtime critic who will be remembered by fans of Alex Bennett's local radio show), Sarah Dougher (indie rock goddess), Dennis McNally (Grateful Dead historian), Howie Klein (free-speech activist and former record label bigshot), Jenny Tatone (music writer), Mike Watt (bass player extraordinaire), my brother Geoff (my brother), and, yes, David Hyman. But despite this all-star list of trusted buddies (and there are more, I just picked the most famous ones) … despite all of them, the truth is, if I was to list my REAL trusted sources, the list would be very short: the A.I. software.

This is how twisted I am: I follow David's suggestion, and let technology serve as a conduit to those I trust, but I end up going around in circles, because what I trust is … technology.

Here are two playlists that demonstrate what is being discussed. The first is an URGE Auto-Mix, based on music I play or add to my library, with fine-tuning parameters to adjust for popularity, familiarity, and release date. This is entirely A.I. driven. The artists include Muddy Waters, the Allman Brothers, Eric Clapton, J. Geils, U2, Brentwood Music gospel, Neil Young, Stevie Ray Vaughan, the Stones, and the Byrds. All but the Brentwood musicians are people I am very familiar with (and I have the familiarity slider set to "not very familiar"). My guess is I'd like maybe ¾ of the tracks, but wouldn't get introduced to anything really new to me.

The second playlist is a "Magic Button" list … the top song this month of people on my Trusted list of MOGgers. The artists include Cursive, the Mamas and the Papas, the Rachels, Professor Longhair, Clipse, Bettie Serveert, Rob Hotchkiss, and Ennio Morricone. My guess is I'd like a lot of this music, but it's hard to say, because a lot of it would be new to me … it's a system that would introduce me to new stuff.

Depending on what I'm looking for (familiar music I will probably like, new-to-me music that is hit-or-miss), both systems "work." But I think the two playlists make a good case for David's idea of "music discovery through people." Accent on "discovery."

Comments