People have strong opinions about music recommendations and almost no information about them.
They're either magic or they're rubbish, depending on the week, and both verdicts often arrive without much sense of what the system is actually trying to do.
Which is a shame, because the basic idea isn't especially mysterious — and understanding it helps explain why recommendations sometimes improve and sometimes slowly narrow.
Recommendations aren't mind-reading.
They're pattern recognition built from your behaviour and the behaviour of listeners whose taste overlaps with yours.
The core idea: taste-shaped neighbours
Recommendation systems often work from a simple and slightly humbling premise:
You are not unique, but you are specific.
There may be nobody with exactly your taste. But there are plenty of listeners whose taste overlaps with yours substantially.
Among the things they enjoy that you haven't heard yet, some are likely to land.
That's one of the central ideas behind collaborative recommendation: find patterns between listeners and content, then use those patterns to identify plausible next things.
Your side
What you actually listen to
Plays, repeats, saves, searches, follows and other signals can help describe the shape of your taste.
Everyone else's side
Patterns of overlap
If listeners who repeatedly enjoy A also tend to enjoy B, that relationship can become useful even if nobody explicitly programmed it.
The interesting part is that the system doesn't necessarily need to understand why two artists fit together in the way a critic would.
It can discover that people who play one also tend to play the other.
It's often pattern before explanation — which is why recommendation systems can occasionally make brilliant connections and occasionally very strange ones.
What it's actually watching
Different services weigh signals differently, but these are the kinds of behaviours recommendation systems can learn from.
1
What you play
A play says something. How long you stay with it can say more. Finishing a track and abandoning one almost immediately are very different behaviours.
2
What you save
Adding something to a playlist or Library is a deliberate action, which can be a clearer expression of interest than a passive listen.
3
What you return to
Playing something once can mean curiosity. Returning weeks later suggests a more durable relationship.
4
What you follow
Following an artist is explicit rather than inferred. You're directly telling the service that this person matters to you.
5
Context and patterns
Some systems can use broader patterns such as when certain kinds of music are played, although the exact signals and weighting vary by service.
6
What you search for
Search is useful because it's direct intent: you went looking for something rather than merely accepting what was already in front of you.
What it can't know
This is the more useful half, because it explains a lot of recommendation frustration.
Why you played something
Forty plays could mean obsession, a child demanding the same song, a party, or somebody else using your account.
Whether silence means love
Not skipping a song doesn't automatically mean you adore it. Sometimes you simply couldn't be bothered to reach for the phone.
What happened in your life
Music attaches itself to memories and circumstances. The listening data doesn't automatically understand those associations.
Who you're about to become
A system predicting from your past has a harder time with something completely outside that past — even if you'd love it the moment you heard it.
That final limitation is the argument for keeping something less personalised in your listening life.
Live radio, editorial programming, a friend's recommendation, or anything else chosen without reference to your listening history can introduce a direction the recommendation system would have had no reason to try.
Personalisation explores around your existing taste.
Unpersonalised discovery can occasionally drop you somewhere completely different.
Why memory length matters more than you'd think
A recommendation system needs some picture of your listening history.
How much history it considers changes what kind of picture it gets.
Shorter window
More sensitive to now
Recent obsessions matter more. Temporary phases can have a larger effect, and older favourites can fade from view sooner.
Longer window
More of the long-term shape
Seasonal habits, older favourites and artists you return to after long gaps have more opportunity to remain part of the picture.
A short window can know your current mood very well and your longer-term taste less well.
It may not know that you return to the same albums every autumn, or that an artist you haven't played in six months is still one of your favourites.
A longer window can capture more of that shape.
In Zene
3 months on Free.
12 months on Premium.
Premium gives the recommendation system a longer history to work from, which can make it less reactive to a short-lived listening phase.
The cold-start problem
A new account knows very little about you.
That's unavoidable, and it's one reason changing services can feel disappointing at first: you're comparing a platform that may have years of behavioural history with one that's only seen a few days.
Two things can speed that process up dramatically.
Import your playlists
A playlist can represent years of deliberate choices. Bringing those collections across gives a new service far more context than starting from an empty Library.
Then just listen normally
Don't turn listening into a training exercise. Search, play, save and skip the way you genuinely would. Real behaviour is the useful input.
A new recommendation system doesn't need you to perform your taste.
It needs enough normal listening to start seeing it.
How to make yours better
Let things play when you actually like them
Constantly skipping after a few seconds creates very different behaviour from actually spending time with the music you enjoy.
Save things deliberately
A playlist or Library save is useful because it's an intentional statement rather than passive exposure.
Play what you genuinely love
If your favourite albums never appear in your recent listening, the system has less evidence that they're still important to you.
Follow artists explicitly
Following is especially useful for favourites that matter to you even when you haven't happened to play them recently.
Give it some variety
If nearly all of your listening comes from one tiny corner of your taste, recommendations naturally have less evidence about everything else you might enjoy.
Keep one unpersonalised channel open
Live radio, editorial programming or recommendations from other people can introduce things your own listening history would never have generated.
And once one of those outside discoveries lands, playing and saving it becomes new input.
Your personalised recommendations can then start moving into an area they couldn't have reached without that first external push.
The honest summary
Recommendations are a mirror with a slight forward lean.
They show you variations and extensions of what your past behaviour suggests you already care about, and they're genuinely useful at that job.
What they're less suited to is surprising you in a direction that has almost no connection to anything you've previously done.
Let the algorithm know you. Don't let it become your entire musical world.
Use personalised recommendations for refinement, and keep at least one route open for music that arrives from somewhere completely outside your history.
That combination — strong personalisation plus occasional genuine surprise — is much more useful than asking either one to do both jobs.
The Home feed and personalised mixes are available in Zene on Android, iPhone and iPad, and the web. Recommendation history covers 3 months on Free and 12 months on Premium.
Conversation
Comments
Post a Comment