Recommender systems and sampling

The advent of digital media has made artistic content more widely accessible than ever before. For the most part, we can find any song, film, or TV show within minutes. Paradoxically, however, this can have a paralyzing effect: the digital media landscape is massive and ever-changing, and finding the content we want requires approaching this landscape with a particular idea in mind.

Drawing an analogy, the biologist Jakob von Uexküll argued that organisms don’t simply experience the world as passive recipients; instead, underlying perception is a search image, an idea of what they would like to find. This observation has a few interesting consequences for how we think about cognition[1] , but the most relevant here is that what we think of as “reality” is actually a noisy, meaningless morass, and different organisms sample and extract different information from that chaos as needed. But as mentioned above, this sampling process requires active engagement with our environment––organisms must learn to leverage their perceptual capacities in species- or even context-specific ways.

Put simply: you can’t find what you want if you don’t know what you’re looking for.

Recommender systems: the solution?

The modern solution to this problem is a recommender system. If you use Netflix, Amazon, Facebook, or practically any other Internet content-provider, you’ve likely benefitted from a recommender system.

The point of a recommender system is to make predictions about what you, a consumer, will like. Given some set of observations about what a consumer has purchased, rated, clicked on, “liked”, and so on, a recommender system tries to predict what else you will purchase (or like, rate, etc.). One relatively straightforward way to do this is to consider which products are “similar” in some way to the products you’ve already purchased, then recommend those. A slightly more complex approach would be to identify other consumers who have purchased the same or similar products as you, then recommend to you the other, potentially similar products they’ve bought (see Amazon’s “People Also Purchased” section).

But the point of this article isn’t about how recommender systems work. I bring up recommender systems because they’re being used to supplant (or augment, depending on your perspective) a critical component of how we experience the world: essentially, recommender systems try to infer our search image from our previous behavior, then query the chaos of the digital media landscape to find the relevant answer. I also don’t think there’s anything inherently “good” or “bad” about this––it’s clear from their existence that at least some have judged recommender systems to be a kind of necessary guide for our online experience.

Accounting for Taste

Why do we like the things we like? Tom Vanderbilt sets out to answer this question in You May Also Like, reviewing centuries of aesthetic philosophy and more recent advances in psychology and machine learning. The book is full of nuggets about human preferences, but what stuck with me was the idea that our preferences, and indeed our enjoyment of or displeasure with particular experiences, are not formed solely from the “thing in itself”. Our subjective experience of anything––music, food, wine––is often the product of our expectations of that thing, our beliefs about that thing, the attitude of our social group towards that thing, and of course, the thing itself.


For example, you might think that a particular wine tastes better if you are told that it’s been aged for half a century; you might be more inclined to like a particular song if your romantic partner and all of your friends love that song[2]. Taste is also associated with cultural capital, so some of us might put in more effort to enjoy certain things, as a way to both acquire and broadcast that cultural clout.

Recommender systems try to learn our preferences and predict what else we will like. Each of us becomes a point in some multi-dimensional “preference space”, comprised of preferences we’ve demonstrated in the past. But if our enjoyment of a thing is formed by nebulous features like social class and prior beliefs––and if the ultimate aim of a recommender system is to replace, or augment, our search image on the web––then should recommender systems attempt to estimate these features as well, in the hope of building more accurate predictive models?


[1] For example, the fact that two different organisms (or even two different people) can “see” the same thing but have radically different experiences of it. Uexküll argued that these different experiences were the product of different Umwelten––essentially “points of view”.

[2] The more contrarian of us might feel less inclined to like that song, though this still reinforces the same point––our like or dislike of something is swayed by how we think others feel about it.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s