Pandora is in a loop !!!!

i have been using pandora for quite some time now. a few weeks ago i decided to create a new station ... and decided to base this station on the track 'hands of time' by groove armada. the songs that pandora threw up were awesome ... and i rated them. some of the recommendations were not upto the mark, and i asked pandora to never play them again.

after about a week of doing this, the station now plays only a handful of songs !!! its almost like i have a playlist on winamp and i am iterating through it at random. occasionally a new song does come up, but the tracklist is pretty much static now for this station.

i must add that i have used all possible rating methods for refining this station. thumbs up/dowm, the 'guide us' pop up menu options ...

what i infer from this? if u spend enough time refining your choices ... u can train pandora algorithm to be very sensitive to your preferences ... and ofcourse the algorithm probably goes into a loop or something ... its almost like you are restricting its preferences and refining its options to one particular line of intrest. for a lack of better word, i would call the pandora system 'breakable' ... not predictable ... but it can be trained.
had a long week so i am going to sign off now ... if you want to see proof for the above just shoot me an email (zarthos@gmail.com) with your pandora account name and i will share the station with you.

Observation: Moods

Not related to last.fm, but i have noticed how moods influence my music tastes. sometimes it is very evident as to how my mood is influencing my taste, but the other day i had a very intresting experience. working at my research lab, i realized i was getting tired of same tracks being churned out by winamp (note: i need get some new music). i tuned into some of the online radio stations and coudnt find anything intresting.

i usually goto pandora.com, cause 9/10 times i always have a good expereince with it. but pandora too disspaointed me. i then fired up last.fm ... got some decent tracks for the first few minutes, but after a few hours i decided that today was just not the day to listen to music !!!

then it dawned on me ... all these music recommenders and radio stations were playing the music i like ... casue i remember rating them and listening to them over and over like a million times. then why today was i finding it do difficult to accept the same music?

then it got me thinking ... we all know moods affect our taste ... but what if i am in a 'i-dont-like-anything' mood, and that day happens to be the day i rate music at pandora or last.fm. what i rate and what preferences i select will definately not appeal to my musical taste-buds later.

should recommender systems consider moods to be a deciding factor when recommending music, and should there be systems in place to prevent users from completely screwing up thier preferences and ratings for their recommender system profile? intresting questions and i am sure someone somewhere is working on this ...

oh btw ... i finally did hit some success with a bunch of rock-house remixes ... funky for my taste lol ! (check coldplay.com ... hit 2003 on their timeline ... nice royskopp remix on there)

Analysis for Feb 14 - Feb 21

Before I begin to summarize my analysis for this week, my findings for the past few weeks:
1. a. Audioscrobbler generates recommendations on a weekly basis, which I think is influenced by the listening habits of my 'neighbours' and of course, myself. The popularity of an artist or track helps it to climb or fall in the recommendations list.
1. b. Audioscrobbler recommends audio tracks, albums, artists, events (concerts), videos and even users (neighbors).
2. I have asked a few people to be my friends, but seems like people don’t like me a lot over here lol ... I don't have any friends as of now, but I hope I do get some soon, which will enable me to study the correlation between their listening patterns and my recommendations.
3. I have a list of neighbors, who are ppl who listen to music similar to what I have been listening to ... last.fm calls them recommended users.

Found an interesting piece of information in the FAQ section:
"The list of artists which you may see on an artist page as being "similar" is based exactly on our user's listening habits. If a lot of users listen to Artist X, but also Artist Y and Z - Y and Z artists will become similar to X.

With a few added constants and other functions thrown into the equation to help make it more accurate, you'll find that's how Similar Artists are made."

I wonder if they use the same technique for track recommendations too...

Now coming to the recommendations for this week, the following were the recommendations for this week:

Artists: linkin park, yardbirds limpbizkit staind were all similar to the tracks I have been listening to in the past few weeks on the last.fm radio(new age rock/metal)
police, jimi hendrix, franz ferdinand were all similar to the tracks I have been scrobbling in the past few weeks. thus we know for sure that last.fm recommendations are influenced by both what I listen to on their radio and on my winamp player, with the scrobbling feature enabled.

music: unfortunately, the tracks thrown up by the recommender engine were not the ones I was looking for. but then I decided to check out my neighbors, and their music was a really appealing to me. last.fm lets you listen to an assortment of tracks your neighbors have been listening too, and this mix of old-age and new-age rock and metal proved an awesome experience. I hope my own recommendations begin to reflect my neighbors taste soon.

Links:
1. to view my recommendations ... log into last.fm with the username zarthos@gmail.com and password 'zarthos'
2. my last.fm profile page

Analysis for Feb 7 - Feb 13

This is the first week when last.fm threw up some recommendations from its recommender engine (Audioscrobbler). To summarize my efforts till now, I have setup last.fm clients on multiple machines and have been feeding my music preferences to Audioscrobbler. After playing close to 550+ songs and joining a music group on the network and adding some friends to my network, I have finally kick started my recommendation profile on last.fm

So what were results of all this? Recommendations … lots of them!!! Some very interesting aspects of these recommendations were:
1. the recommendations are classified based on where they have been derived from (group recommendations, friend recommendations, last.fm recommendations).

2. it looks like all these recommendations are tied up by Audioscrobbler to my listening preferences, but the recommendations vary very widely depending on the sources.

3. Recommendations from friends reflect their taste and I found most of these recommendations to be far from what I was looking for. But this is a social networking based technology, and listening to other people sometimes does open up interesting choices.

4. The recommendations are also classified based on which week they were produced. This is amazing, cause my taste can change from week to week, and so can that of my friends and groups. Thus the engine does not attempt to judge your preferences based on cumulative knowledge, but rather focuses on a weekly preferences and changes in listening habits too, though it does provide an option to get a list of recommendations based on cumulative listening habits.

5. Neighborhood of friends!!! Yes we can actually see all the people who are listening to music that u have in your profile. This neighborhood helps to showcase 2 things:
a. Popularity of the music you are listening to.
b. Potential friends who you can add to your list.

6. There is also a recommendation list for albums and concerts and other music related events, though I have not analyzed these recommendations in detail.

What I plan to do next.
1. Explore the tagging functionality and determine its influence on the recommendations generated by Audioscrobbler.
2. Can I make the system unlearn my preferences … which is very easy to do in Pandora?

There is a note on how recommendations are generated and tagging in the help section, which is where I plan to begin my hunt for more information. Also I plan to profile some of the recommendations to determine a pattern, if any, in the recommendations being generated.

Links:

1. to view my recommendations ... log into last.fm with the username zarthos@gmail.com and password 'zarthos'

2. my last.fm profile page

Finally ... some recommendations !!!

Finally, after waiting for what seemed like almost an eternity, last.fm is throwing up some recommendations. it seems their website is plagued by a host of issues including artist name mismatch and other technical bugs, which prevents me from exploring some of their advanced features. nonetheless, i will be reviewing my recommendations soon and try to find a pattern, if any, in them.

keep an eye on this page for more updates on the last.fm recommendations.

Observation: Track length

I made very intrestesting discovery today as i was trying to play some of the tracks in my playlist.
if the track lenght is less than 30 seconds, last.fm radio player refuses to scrobble the track. this might provide the answer to Dr. Wilson's question in class: how does last.fm know if u like a track or not? just because u play a track for 5 senconds need not signfy the fact that you actually like the track.

Feeding the scrobbler

This week, my objective was to setup a account with last.fm and build a user profile by feeding data to the recommender system. I hooked up winamp with the recommendation engine by installing the required plug-in. I then queued close to 500 songs on winamp. While adding songs to the playlist, I made sure I chose songs from similiar genres. But at the same time, I also had to make sure I wasn’t biased towards anyone particular artist. Some of the artists and songs I had on my playlist were:

500 George Thorogood - Move It On Over
499 Bob Dylan - Rainy Day Women #12 & 35
498 Joe Cocker - The Letter
497 Ten Years After - I'd Love To Change The World
496 Kingsmen - Louie Louie
495 Humble Pie - Thirty Days In The Hole
494 Beatles - I Should Have Known Better
493 Bob Seger - Old Time Rock And Roll
492 Santana - Evil Ways
491 Robert Palmer - Addicted To Love
490 Lovin' Spoonful - Summer In The City
489 David Bowie - Rebel Rebel
488 Amboy Dukes - Journey To The Center Of The Mind
487 Ozzy Osbourne - Crazy Train
486 Sly/Family Stone - Dance To The Music
485 Creedence Clearwater Revival - Suzie Q
484 Warren Zevon - Werewolves of London
483 Eric Clapton - Forever Man 482 Hollies - Bus Stop
481 Sugarloaf - Green-Eyed Lady
480 Peter Gabriel - Sledgehammer
479 Rolling Stones - Let It Bleed
478 Tom Petty - Free Fallin' 2
477 Neil Young - Southern Man
476 Crosby, Stills Nash - Helplessly Hoping
475 Chambers Brothers - Time Has Come Today
474 Santana - Black Magic Woman
473 Eagles - One Of These Nights
472 Canned Heat - On The Road Again
471 Dr. John - Right Place, Wrong Time

I then let winamp do its thing, and let it play for upto 48 hours on a roll (a song averages 3 minutes, so for 500 tracks the total playtime would be close to 25 hours). Ofcourse, I did have other playlists pumping in music from artists like audio slave, Coldplay, Perl jam, dream theatre, etc.
Note: when I am playing these tracks, I am not able to rate them. So audioscrobbler has no idea as to weather I actually like the tack or not. It assumes that just because I had the track on my playlist and I am playing it, I like it. Also, from my initial observations, it seems audioscrobbler send the name of the song to the last.fm server as soon as I hit the play button (or as it comes up next on the playlist)

Now, I have all this information fed to the recommendation engine. Is it enough to get the recommendations I need? What will be the quality of the recommendations? What other features of last.fm leverages the recommender system funcitonalities? To answer all these questions, I have to wait a while, or atlesat that’s what last.fm says. Lat.fm is calculating my recommendations, and in this time, I am going to feed more music to it!!! I also plan to rate some the tracks I listen to on the last.fm radio player. All this will hopefully give audioscrobbler enough data to make a good set of recommendations for me.