A Sense of Doubt blog post #1610 - How does the Pandora Algorithm work?
I need a quick share. I researched this years ago and never posted it.
The song selection algorithm was a bit mysterious in its choices.
There seems to be somewhat of a difference of opinion in Quora users on the way the algorithm picks music in your station themed by your initial artist or genre or time period selection. T he most compelling answers seem to be that the algorithm matches by musical attributes rather than user ratings or your own preferences, though some believe the latter and it may be true.
I sense that these posts are just the beginning of the conversation and that I have more exploring to do to reach a definitive understanding.
https://www.quora.com/How-does-the-Pandora-algorithm-work
Joey Flores, Capitalist Hippie and VP of Biz Dev at flexReceipts
Sorry to disagree with Sam but it is definitely an algorithm. It's an algorithm based on musical traits that are entered into the system by music experts.
From what I understand...
The music is assigned attributes such as "heavy guitar riffs", "metal influences", etc. I believe they have somewhere around 4000 of them. The list is on Wikipedia and is a bit of a disaster, with a lot of the same terms used multiple times differently, which is probably one reason why classifying a song takes 15+ minutes.
When you type in that you like Metallica, it looks at the most frequent traits in their songs and then searches for songs by other artists with similar qualities. It doesn't always play songs with the absolute most similar qualities. It plays 4 song batches with one set of similarities, then plays 4 with other similarities. A good example is that it might play 4 acoustic songs, happy songs, etc.
Every time you thumbs up a track, it adds it to the artists your station is based on and also plays that song a bit more often. Similarly, when you thumbs down a track, it takes it out of rotation and skips to the next 4 song batch (assuming you may not be in the mood for that style right then). If you thumbs down two songs by the same artists without ever thumbing one up, supposedly they will never play that artist again. Also, each station is independent of others. Thumbing a song in one channel does not affect the preferences of the others.
So, assuming you actively rate songs, over time, your station will evolve, play a wider selection of music and, theoretically, better represent your taste. If you ever want to tweak a station, you can go into your preferences and remove the thumb ratings from various songs, if you think you can tell which ones started resulting in music you didn't like.
From what I understand...
The music is assigned attributes such as "heavy guitar riffs", "metal influences", etc. I believe they have somewhere around 4000 of them. The list is on Wikipedia and is a bit of a disaster, with a lot of the same terms used multiple times differently, which is probably one reason why classifying a song takes 15+ minutes.
When you type in that you like Metallica, it looks at the most frequent traits in their songs and then searches for songs by other artists with similar qualities. It doesn't always play songs with the absolute most similar qualities. It plays 4 song batches with one set of similarities, then plays 4 with other similarities. A good example is that it might play 4 acoustic songs, happy songs, etc.
Every time you thumbs up a track, it adds it to the artists your station is based on and also plays that song a bit more often. Similarly, when you thumbs down a track, it takes it out of rotation and skips to the next 4 song batch (assuming you may not be in the mood for that style right then). If you thumbs down two songs by the same artists without ever thumbing one up, supposedly they will never play that artist again. Also, each station is independent of others. Thumbing a song in one channel does not affect the preferences of the others.
So, assuming you actively rate songs, over time, your station will evolve, play a wider selection of music and, theoretically, better represent your taste. If you ever want to tweak a station, you can go into your preferences and remove the thumb ratings from various songs, if you think you can tell which ones started resulting in music you didn't like.
Pandora relies on a Music Genome that consists of 400 musical attributes covering the qualities of melody, harmony, rhythm, form, composition and lyrics. It's a project that began in January 2000 and took 30 experts in music theory five years to complete. The Genome is based on an intricate analysis by actual humans (about 20 to 30 minutes per four-minute song) of the music of 10,000 artists from the past 100 years. The analysis of new music continues every day since Pandora's online launch in August 2005. As of May 2006, the Genome's music library contains 400,000 analyzed songs from 20,000 contemporary artists. You won't find Latin or classical yet: Pandora is in the process of developing a specialized Latin music Genome and is still deep in thought about how to approach the world of classical composition.
so in mathematical terms (S, T)=sqrt [(for i=1 to n)Σ(si−ti)ˆ2]
Where the difference of each attribute is calculated to show the most applicable music.
That's the basic formula.
Kayaker - 03-29-2016
I use Pandora Radio, paying monthly so I don't have to hear ads. I've played around and created a few stations I really enjoy, but I have a few questions maybe someone could help me with.
1) Strategy.
I "like" (thumbsup) stuff that I like. A friend told me that is not a good thing to do, as they'll overplay any song that you thumbup. He, instead, assumes he will like anything they play on a station. He never thumbsup, instead he thumbsdown anything he hates. That way, per him, each station has more variety. Is this correct?
2) Thumbprint
OK, a relatively new offering is the Thumbprint Station, which is made up of all the thumbsups you've given on all your stations, plus related songs. I love this station. Obviously, my friend who never thumbsups does not have a thumbprint station. His loss.
I occasionally find that the thumbprint station gets stuck in a certain genre. Anyone else have this problem? I just switch stations, come back, and all is well.
3) Algorithm
So, I have a Dusty Springfield station. Please don't hate me. I have consistently thumbsdowned anything that doesn't fit perfectly and thumbsupped everything that embodies Dusty Springfield's work. For those times when I need some Dusty Springfield.
It works. I love the station.
But one day I felt like some Dusty Springfield. After a few very fitting songs there was a song that absolutely did not fit. I was surprised. Then I looked to see who the band was. Buffalo Springfield. Was Pandora trying to be cute, or does the algorithm work like that?
4) Interesting weirdness
I was playing around with a comedy station. Just to see what would happen, I intentionally thumbed down any non-Caucasian or female comedian, curious to see how the algorithm would respond.
In a short time I created a horrid comedy station that was all red-neck southern humor. Interesting. I deleted the station.
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
- Bloggery committed by chris tower - 1907.18 - 10:10
- Days ago = 1475 days ago
- New note - On 1807.06, I ceased daily transmission of my Hey Mom feature after three years of daily conversations. I plan to continue Hey Mom posts at least twice per week but will continue to post the days since ("Days Ago") count on my blog each day. The blog entry numbering in the title has changed to reflect total Sense of Doubt posts since I began the blog on 0705.04, which include Hey Mom posts, Daily Bowie posts, and Sense of Doubt posts. Hey Mom posts will still be numbered sequentially. New Hey Mom posts will use the same format as all the other Hey Mom posts; all other posts will feature this format seen here.
No comments:
Post a Comment