The Perception of 'Random' on the iPod

Robaato writes "Stephen Levy writes in the Guardian about the perception of randomness, or the lack thereof, on an iPod set to shuffle." From the article: "My first iPod loved Steely Dan. So do I. But not as much as my iPod did.... I didn't keep track of every song that played every time I shuffled my tunes, but after a while I would keep a sharp ear out for what I came to call the LTBSD (Length of Time Before Steely Dan) Factor. The LTBSD Factor was always perplexingly short." My first iPod shuffle refused to let me delete (sigh) Weird Al's Polkamon off of the flash memory.

OCD

[eldavojohn]

Or - and here is the nub of an issue that would consume me for more than a year - was the shuffle function, meant to mix up my music collection in a random fashion, actually not random at all?

There there, Mr. Levy, we'll get you all the randomness you want [wikipedia.org]. In fact, we have a special place filled with randomness and padded white walls! You're going to like it there.

You know, instead of wasting your interviewee's time, you could have installed a five song list on your iPod and set it to shuffle. You'd have to carefully mark down the track number being played and listen to it for 100 songs. Do this a few times and make sure you're very methodical about what you do. Wipe the iPod, put five songs on it in order and then listen to a hundred songs "randomly." If you start to see a pattern developing or one song is obviously favored over the other, it will begin to show up.

But on the more technical side, they have to seed the random variable with something. Whether or not it's an internal clock, I'm not sure. Either way, they have to derive a random number and it's possible that their seed isn't good enough or has too few states or is prone to being seeded at the same state, etc. Based on this information, I hate to break it to you but it is very hard to be truly random.

People are Pattern seekers

[BigDiz]

Humans innately seek patterns in things that are random. That's why so many people wear smelly socks because they think they're lucky. Once you identify a supposed "pattern" i.e. non-randomness, you're going to keep noticing instances that fit that pattern, and ignore instances that do not. This is deeply ingrained.

Think about it, if you're at the roulette table and black has come up four times in a row, how likely are you to bet black? Most people would bet red, because, I mean hey, there's got to be a pattern. But (as I'm sure you all can understand) black has the same probability of occurring again as red does.

People have had this complaint about all sorts of playlist randomizers (not just iPod), it's just people seeing what isn't there.

Two random modes

[daeg]

Apple should add another random play mode -- one that acts as it does now, and the other mode that grants every song an equal play count. The only thing that would be random is which order. This way users that have a confirmation bias of their iPod favoring certain songs can no longer be paranoid of Apple conspiracies to promote the songs of {{ artist }} or {{ record_label }}.

Re:Never true randomness

[Reaperducer]

I have one but only ever use it on long journeys and no I don't have DRM'd tracks so I didn't care about online music purchases. The ipod just happened to be the one that worked the best (scrollwheel is nice and quick) and having a mac I knew it'd work well.

It's a shame how people on Slashdot aren't allowed to just like iPods -- they always feel pressured to justify the purchase.

"Best tool for the job" isn't good enough. You have to be different. But only in a pro-Linux anti-iPod sort of way. Any other kind of being different gets you modded troll or flamebait.

Re:And Zonk dupes himself... again...

[Anonymous Coward]

I remembered it, and I didn't get paid to post it.

Re:Bias - hmm

[sreekotay]

Not sure its so much confirmation bias (alone at least), as it is that the odds of NOT playing a song from the same artists over the next X songs shrinks more rapidly than intuition suggests. That is, for example the odds of NOT having a run of X heads or Y tails when flipping Z coins is very, very small.

The article mentions the "how many people does it take to get to a shared birthday thing" - and the point there is that its not that it takes 40 people to get to one with a SPECIFIC birthday but only 40 or so to find two that SHARE a birthday.
-----
graphically speaking [kotay.com]

Re:People are Pattern seekers

[Anonymous Coward]

Exactly, if you show people two series of dots, one really random and one evenly distributed, but not regular and ask them which is more random, they will say the evenly distributed one, becasue the random one has what we see as obvious patterns in it. So, I think what people want is evenly distributed (but not regular) mixing of songs, not true randomness.

Humans and dictionaries define random differently

[jonadab]

Except for mathematicians and programmers, most think of "random" in a *very* different way from its technical definition. To most humans, saying that a particular sequence is "random" means *guaranteeing* certain things about it. Among them: the same element does not occur back-to-back, EVER, even if there are only a few elements total to choose from. Even more, if there are more than about half a dozen elements, the same element never occurs twice within about five positions. (So if you've got songs 1 through 7 on your iPod, and the first seven played are 5, 3, 7, 2, 4, 1, 6, then the next one has to be 5 or 3, or _maybe_ 7, or it doesn't seem "random" to most people. Yet, the order can't be the same every time through, either.) No element occurs substantially more often than any other element, even over the short term. If the elements have a natural order (e.g., alphabetical), then no three elements that are adjascent in that order can ever occur together in that order, nor should they occur together in the reverse order. (This gets particularly difficult to guarantee when the elements have more than one natural order, e.g., if the elements are people, you can't have three of them in a row by either name or age, or people notice and decide that the order is not random.) Even worse, if the elements can all be categorized into a small number of categories (e.g., by gender), you can't have "too many" from one category in a row. (How many is too many depends on the ratio, but if half of the elements are male and half female, having four of either in a row will make people cry foul, the order is not "random".) If certain elements stand out from the others in some significant way, they can neither occur first nor last. (For instance, if test questions are being drawn from a question bank, neither the easiest nor the hardest question should be first or last; if it is, people will say the order was not random.)

I could go on and on, but what it really amounts to is that when most people say "random" they mean "carefully arranged in a thoroughly mixed-up order". This is almost the *opposite* of what a mathematician or computer programmer thinks the word "random" means.

For this reason, when describing a mathematically-random sequence to an end user, I never EVER use the word "random". I generally call it something like "arbitrary" or "unpredictable". This greatly reduces complaints.

Now, as far as song frequency, I like to rate my tracks on a scale of 1-10, and rig my playlist so that anything under a 6 never plays unless I specifically select it, tracks rated 7 play twice as often as those rated 6, and the frequency keeps going up the higher my rating is. (I only have eight tracks rated as a 10, and they're all things I don't mind hearing back-to-back.) Then if I find a track is playing more often than I like, I figure I rated it too high and cut back its rating.

Re:Bias - hmm

[Anonymous Coward]

Yes, that's a good point (and true). However, the original article is talking about one particular artist being played more often than expected. You're exactly right in saying that the gap between songs from the same artist in a shuffled list is shorter than most people would expect. The article author, however, is probably wrong in saying that the gap between Steely Dan songs is less than the gap between another artist.

Re:Mine loves Chevelle

[philipgar]

if the iPod/iTunes sensed higher played songs and played them more, it would actually cause a huge disaster, depending on how it was implemented. A simple implementation would increase the likelihood that a song gets played again by a fixed percentage. Under a situation like this, no matter how small of a percentage the song increases it's likelihood of being played by, and regardless of how many songs you have on the device, over enough time it will reach the point where it will effectively only play one song over and over 99.9999999% of the time. It's the problem with positive reinforcement as the song will become more likely to be played the more it's played, and once one song gets "lucky enough" it will be promoted just high enough to start dominating, and from there it quickly runs away. I ran a simulation similar to this not too long ago as I was thinking about it.

Now, if you scale the likelihood of a song being played by a fixed scalar value, this runaway effect won't happen, but that isn't as much fun to do, and would mean overtime the scaling would be meaningless as each song has a high enough value that the scalar/value is effectively zero (or on a floating point number could be zero).

Phil

Re:Truly Random

[Jerf]

It's not cost-effective to be "truly random" with an iPod Shuffle's hardware budget, but who cares. It's not at all hard to convince a pattern-seeking human that it's not "random", because our pattern-seeking human thinks "random" means "play everything once, then start over with a fresh list", which is anything but random.

I just ran a simulation here of 100 songs, randomly chosen until all 100 songs had been selected once, and ran it 1000 times. On average, it took 523 choices to exhaust the list due to repeats, and on average, the most-played song over that time period had been played 11.7 times before the last unplayed song was finally played.

Joe User will of course assume that the iPod "likes" the song it plays 11.7 times and "hates" the song it only "begrudgingly" gets around to playing 500+ play-throughs later, but that's true randomness.... and one of my pet peeves with MP3 programs.

Humans very rarely want true randomness. When we claim to want randomness in our shuffle, we actually want a somewhat more sophisticated algorithm. But programmers hear "random" and, well, the easiest thing to do is just call the "randint" function... I did once write a non-random shuffler for a thoroughly-non-mainstream music program and I thought it worked out pretty well, but I've never been able to interest anyone in the idea, because it seems like a lot of implementation work vs. "randint".

import random
from operator import add
 
amounts = []
maxes = []
 
def average(seq):
    total = reduce(add, seq) + 0.0
    return total/len(seq)
 
for x in range(1000):
    choices = {}
 
    i = 0
    while 1:
        i += 1
        next = random.randint(0, 99)
        if not choices.has_key(next):
            choices[next] = 1
        else:
            choices[next] += 1
        if len(choices.keys()) == 100:
            break
 
    amounts.append(i)
    values = choices.values()
    values.sort(lambda a, b: cmp(b, a))
    maxes.append(values[0])
 
print "Average choices: %s" % average(amounts)
print "Average max: %s" % average(maxes)

Consider this program on par with any other Slashdot post; it may have typos, it may even be flat wrong, it certainly wasn't engineered for use in an enterprise-class environment nor analysed for how to make it run faster or with fewer keystrokes, etc. It wasn't composed in a browser text-box but only slightly more care was poured into it.

As I like to say, "random is blotchy". Smoothness is a dead giveaway that a process isn't random. Randomness produces bell curves, not uniform distributions.

Re:Bias

[GeffDE]

And I want iTunes to make me a continental breakfast every day by 7 sharp and then clean my house and prepare dinner for when I arrive home.

iTunes is a jukebox program, not your own robotic DJ. iTunes does give the option to weight the play-order based on how long it's been since a song played. Go to your library, turn off "shuffle mode," sort the library by Date Played and play the songs you heard last. In order. There you go. 100% weighting on date played. iTunes doesn't make a note in its database if you've skipped a song because it doesn't include one of those fucking mood rings so it knows what color you are when you want to listen to Kelly Clarkson vs what color you are when you want to listen to Celine Dion. And how can iTunes tell when you heard it too recently? Maybe you just heard a new song, and like it a lot, so you want to hear it a lot. Does iTunes know that? Is iTunes supposed to have a T3 line to Ms. Cleo's brain?

Lastly, iTunes sucks down quite a few cycles even during normal playback. And when it sucks down too many, it bogs down the system. Now, when I have iTunes on random mode, it can also be referred to as "I was iTunes to provide ambient noise for while I am working on other things." And when iTunes starts interefering with that, I get cranky. I like that iTunes does what it is supposed to do, and doesn't try to do everything. If you want something that will wipe your ass for you, find a maid, not a jukebox program.

Re:Bias

[GregWebb]

Agreed. Evenly weighted (or biased according to preferences) shuffle is far more useful than random.

In my last job, I listened to MP3s all day while working. I'd got RoboDJ feeding in the playlist, and Audioscrobbler logging the results.

Which turned out - that, remarkably consistently, Iron Maiden, Placebo and System Of A Down (typically 'Aerials' from Toxicity) got played more than their weightings justified, while Deep Purple measurably less than theirs. Including compensating for number of tracks held.

It may have been random, it may have been provable that each run had even distribution, but in the runs I was doing there it would have been far more useful if the computer had logged what it had already played, and tried to keep a track from getting too far outside its stated range. Less random, yes - less what I asked the computer for, definitely not.

Re:And Zonk dupes himself... again...

[mikesd81]

Well, there is times I wanted to post an article but wasn't sure if it was posted yet and I searched for it. Surely he can too?