Program Directors > Recommenders

The endless debate on the best approach for providing music discovery experiences continues rage. This continued with a SxSW panel put on by Paul Lamere (from The Echo Nest) and Anthony Volodkin (from The Hype Machine). For some background, here is their presentation:

Help! My iPod thinks I'm emo.

View more presentations from Paul Lamere.

Additional notes from Anthony are here: http://fascinated.fm/post/89782283
And notes from Paul can be found here: http://musicmachinery.com/2009/03/26/help-my-ipod-thinks-im-emo-part-1/

My take? There are many ways to classify these different approaches, and some can be considered subsets of others. But, in short there are more approaches than most can shake a stick at... with some of the more notable examples being (as I see it):

• "Musicologist"/Expert Analysis - a hundred people with headphones on categorize songs on a few hundred different attributes (e.g. minor chords, upbeat, heavy piano)
• Content-Based/Waveform Analysis - a computer looks at the sonic attributes of a song (e.g. tempo, harmonic range, etc) and then looks for other songs that posses similar attributes
• Collaborative Filtering - people who listen to/buy/highly rate song X, also have a high correlation to buying/listening to/highly rating song Y
• Editorial (e.g Bloggers, Pitchfork, Radio) - some one broadcasts their opinion... "hey, these guys sound like so and so and they are good. I give them 5 stars".
• Semantic (e.g. tag clouds) - by matching tags, genres and sentiment keywords (either manually input by users or extracted from web articles)
  
• Curator (e.g. playlist sharing) - community generated "top song" and "just listened" charts, curated playlists, TV/movie soundtracks. These could be actively curated lists of songs (e.g. playlists) or passively programmed by the curator (e.g. "just listened").
• Shuffle - just a random selection of tracks from a fixed set of songs
• Biographical - this artist you like was influence by this other artist or was also in this other band
  
• Friend-to-Friend - the most common, one friend tells another... "hey, have you heard X? You'd like them".

For each, you generally have two sources of information which to base your recommendations on:

• Implicit - input/feedback is collected passively/automatically (e.g. listening history)
  
• Explicit - input/feedback is solicited and manually provided (e.g. user ratings)
  

Some of the approaches above are exclusive to one type of data input (e.g. "editorial" is explicit feedback only), but most incorporate both.

And, as we speak there are hundreds/thousands of Music Information Retrieval PhD candidates figuring out new approaches, there are new spins (and terminology) on the existing methodologies, and companies combining two or more of the approaches above. Having spent some time in the discovery/recommender space myself, my take is this...

There is no magic bullet.

We all discover content everywhere, and from every context.... friends, TV soundtracks, recommendation engines, radio, etc. In my opinion there is no such thing as a "better recommendation".... there is only a "good enough" (in that what I discovered was satisfying) or a "bad" recommendation. What is required in both cases is transparency as to why it was recommended. I say often when it comes to recommendations... "the why is usually more interesting than the what". That is because the "why" helps user identify the sources that they trust, can relate to, and can be turned to again in the future for inspiration. Sometimes that source may be a machine, sometimes it may be an editor, sometimes it may be a chart. They all have their value, and they all serve the same purpose.. to be one my personal "program directors".

I had a conversation this morning with J T. Ramsay where he told me: "[music] discovery is one of the biggest fallacies of all time. If people were so amped on discovery, radio would have been formatted differently." And I have to admit, I don't disagree.

What consumers have continually asked for is "programming", whether is be playlist builders, charts, "just played" data, recommendation engines, DJs, mixtapes or CDs. Basically, it just comes down to "don't force me to make a decision every 4 minutes about what I'm going to hear next". As music services continue to move to small UI devices where search and destroy is not a viable use case (mobile, car stereos, etc.), programming becomes even more essential.

The title of this post "Program Director > Recommenders" is not meant to imply that they are different and that one is better.... what I mean is that one is a superset of the other. Trying to pit them in a battle with each other is akin to asking... "which is better when you are hungry, a sandwich, a hot dog, or an apple?". The answer for me is, "they will all do the trick... it's just a matter of what temporal mood I'm in weighed against how much effort I have to expend to get each ."

Related articles by Zemanta

• The Echo Nest PromoBot (stylman.tumblr.com)
• Music Bloggers and The Hype Machine Music Blog Zeitgesit (evolvor.com)
• Thesixtyone Is Building a Digg For Indie Music (techcrunch.com)
• Hype Machine: Just let music bloggers influence you! (fridayfivesatsix.com)
• Last FM Boffin (rubbishcorp.com)
• Ten Social Music Sites that Let You Listen, Share and Enjoy Music (rotorblog.com)
• The Music Industry's New Internet Problem (businessweek.com)
• The Lonesome collaborative playlist - Spotify (lonesomemusic.blogspot.com)
• MOG Has Created The Ultimate Streaming Music App; Too Bad It May Never Launch (techcrunch.com)

RecSys 2007

I spent a few days in Minneapolis last week attending the Recommender Systems Conference. The conference is in its 2nd year and brings together academics and industry to discuss the future, challenges and opportunities for recommender systems. This year there are about 110 participants from 16 countries, including industry representation from Google, Netflix, Amazon, AMG, Digg, AOL, eBay, Unilever, Aggregate Knowledge and MyStrands.

Day 1 featured a very interesting keynote speech from Khrishna Bharat, Principal Scientist from Google, about the history and future of news journalism and the social responsibility we all share in ensuring the continued freedom of speech. He also touched on the process by which Google crawls, clusters, ranks, classifies the most relevant stories in Google News. Followed by some insight into the increased user engagement they were able to realize with the introduction of their personalized news stories. The clickthrough of personalized news stories is indeed higher than on just a blind list of "top stories".

The keynote was followed by a number of academic papers presentations - focused on the hot topics of privacy and trust in collaborative filtering engines. Indeed some very interesting research going on in these fields, and I look forward to seeing what the continued research here bears out in the coming months and years.

After lunch, I was honored to take part in a panel with the focus of "Where should we be investing most in research and practice to increase the value of recommenders?". This was the opportunity for the industry folks like ourselves to provide some insight to the academics about the "real world" issues that we are trying to solve or improve. It was a lively discussion that extended the dialog on recommenders beyond the science and into user experience, consumer value and business models built around them. The panel included:

• Joaquin Delgado, CTO, Lending Club Corp.


• Jason Herskowitz, VP of Consumer Products, MyStrands


• Kartik Hosanagar, Assistant Professor, Wharton School of Business, University of Pennsylvania


• David Jennings, DJ Alchemi LLC


• Zac Johnson, Product Manager, All Media Guide, Inc.

The day closed out with Poster Sessions by the academic community and some very interesting demos, with the lively discussion moving on to dinner and drinks.

The second day presented us with more research papers and another industry session titled "Appraising Recommender Systems" featuring:

• Jennifer Consalvo, Director of Personalization, AOL

• Greg Linden, Founder, Findory, Inc.

• Shail Patel, Platform Leader, Unilever Corporate Research

• Neel Sundaresan, Director, eBay Research Labs

• Tim Vogel, Chief Scientist, Aggregate Knowledge, Inc

All-in-all, the industry folks (myself included) challenged the academics with problems and questions.... not answers. Some of the ones I found more interesting were:

• How conservative should a "good" recommendation be? The pro is the con, in that a conservative recommendation is rarely wrong, but also just as rarely leads to a serendipitous discovery.
• When is a recommendation "good enough"? Where is the point of diminishing returns in further research into the algorithms?
• How do you differentiate based on algorithm? Is it possible, or do companies need to focus on differentiating the experience they present *around* the algorithm?
• Do consumers even want the "best" recommendation, or just the most useful? Greg Linden suggested that if Amazon just recommended Harry Potter to every customer, that would probably be the *best*, but not nearly as useful the consumer as recommending something less obvious.
• How do you present a "story" around a recommendation that makes it interesting enough for a user to invest in?
• Can the industry get behind a standard "taste data" format that enables users to own their preferences and consumption history and seamless share that information with any site they desire without having to train yet another system?

The side-benefit of this trip is that I got to meet a number of "Facebook Friends" in person for the first time - David Jennings, Paul Lamere, Zac Johnson, Oscar Celma and others from the "music 2.0" community. Sorry about the tequila shots guys... not my idea. :-)

Music Recommenders - Which Do You Recommend?

Paul Lamere (Sun Labs) and Oscar Celma (Music Technology Group) have just posted their recent presentation on music recommenders at the International Conference on Music Information Retrieval conference. It is a good read (albeit probably too math-heavy in parts for most) and I recommend (no pun intended) that you guys check out the slides if this is a space that interests you.

:: MUSIC RECOMMENDATION TUTORIAL -- ISMIR 2007 ::: "As the world of online music grows, music recommendation systems become an increasingly important way for music listeners to discover new music. Commercial recommenders such as Last.fm and Pandora have enjoyed commercial and critical success. But how well do these systems really work? How good are the recommendations? How far into the 'long tail' do these recommenders reach? In this tutorial we look at the current state-of-the-art in music recommendation. We examine current commercial and research systems, focusing on the advantages and the disadvantages of the various recommendation strategies. We look at some of the challenges in building music recommenders and we explore some of the ways that Music Information Retrieval (MIR) techniques can be used to improve future recommenders."

It would have been great to see the informal survey extended to track-level recommendations (beyond just the artist level), but it is still very interesting even at this higher level.

On a related note, I will be on an panel next week at the Recommender Systems Conference in Minneapolis. Specifically the panel is:

Friday Afternoon Panel: Where should we be investing most in research and practice to increase value of recommenders?

* Moderator: Todd Beaupre, Yahoo, Inc.
* Joaquin Delgado, CTO, Lending Club Corp.
* Jason Herskowitz, VP of Consumer Products, MyStrands
* Kartik Hosanagar, Assistant Professor, Wharton School of Business, University of Pennsylvania
* David Jennings, DJ Alchemi LLC
* Zac Johnson, Product Manager, All Media Guide, Inc.

What's Up with Amazon's Music Recommendations?!

So I'm just diving into Amazon's new MP3 offering and I came across something strange. When I am at their regular CD store I get different music recommendations than I get at their new Amazon MP3 beta. Supposedly, they are both based on "items I own" (or more specifically, have purchased from Amazon). But, if that were the case shouldn't they be recommending the same things?


Amazon Music Recommendations:
* Lots of Fiona Apple and Peter Gabriel - not really representative of what I like

Amazon MP3 Recommendations:
* Arcade Fire, Ryan Adams, Neutral Milk Hotel, The Hold Steady - these are on the money (although I already own all of the things they are recommending)


These are apparently being calculated off of different data? Why the huge discrepancy?

UPDATE:
Someone suggested that this may be due to the fact that the MP3 library is a subset of their CD inventory and therefore so are the MP3 Recommendations. That was my only real theory as well. I'm not sure what labels Peter Gabriel and Fiona are on, but I can look.

But, what I'm actually more surprised by are the good MP3 recommendations. I haven't bought a CD from Amazon in ages yet they seemed to have some insight into my current listening behavior - it's almost like they are scraping data from some of my social networks, taste APIs (like MyStrands or Last.fm), pulling library info off of my machine, or some how getting more current data of my behavior. Amazon shouldn't know that much about my listening habits, and I highly doubt they could gather such information based on my last purchases of computer accessories and baby toys.


UPDATE #2:
Another suggestion was that the MP3 recommendations were being based on my page viewing history on Amazon. The pages I've viewed at Amazon are only for TVs, ink jet cartridges and toys - I don't browse music there. But, I wonder if any of the album artwork fetchers I use (that pull from Amazon) end up inadvertently sharing my library data back with them (does it appear like I have "viewed" those pages?). If that was true and it didn't have any MP3 purchase behavior to base my MP3 recommendations (and only page views), that could explain it....

One Llama

I haven't had time to review it in detail myself yet, but check out Somewhat Frank's coverage of One Llama...

One Llama For 'Fun' Music Discovery : Somewhat Frank :: web 2.0 ● technology ● life :: blog by Frank Gruber

My first impressions are positive ones. Clean/fun UI, easy to generate playlists and the integration with Rhapsody is really nice, although not very obvious (hint: generate a playlist then click "play all" in the upper right of the page). It generated some pretty good recommendations based on a seed... some obvious some a bit more obscure.

The playlist above was created using "Bridge and Tunnel" by The Honorary Title as the seed. Some of these artists I know, others I don't. Although, the first generation of this playlist included "Raindrops Keep Falling on my Head" by BJ Thomas. That seemed a bit out of left field, but I've learned to love the unexpected in recommendations... they drag you out of your comfort zone and expose some new things every once in a while (although in this case I deleted the song and it was instantly replaced with something else).

My only real complaints is that the whole experience seemed a bit too solitary for my liking... even with the user profiles integrated "post to Facebook" feature. And the other is the embedded widget itself (above) - while is is pretty slick looking - there doesn't seem to be a way to get to the page and/or my profile itself (which is the only place you can leverage the freeplay Rhapsody integration. So, in that case you are stuck with a 30-second sample experience only. Not hard to fix... just a couple of hiccups on an otherwise impressive first release.

ACM Recommender Systems 2007

I have the honor of participating in the upcoming ACM Recommender Systems 2007 Conference as a member of the Industry Program Committee, which at this point also includes:

• Francisco J. Martin (co-Chair), MyStrands
• Todd Beaupre, Yahoo! Inc
• Rafael Camargo, Motorola, Inc
• Maikki Frisk, Nokia Corporation
• Jason Herskowitz, AOL Inc
• Kaushal Kurapati, Ask.com
• Dominic Pride, France Telecom, SA"

To learn more about it, or if you or your company is interested in particpating, you can find more info here.