Number Class
← Back to Explore
Computer Science Artificial Intelligence & Data Vector Embedding Design

Your Recommendations Aren't About You, They're About the Crowd

A 2024 Netflix paper found cosine similarity can be mathematically arbitrary, and item embeddings correlate with popularity at 0.98

S.J. Nam 8 min read
Your Recommendations Aren't About You, They're About the Crowd

The Recommendation That Wouldn't Die

In 2017, an Austrian audio-research team named Arthur Flexer and Jeff Stevens sat down with the visualized nearest-neighbor graph of a real, working music recommender and found something odd: certain songs kept showing up in everyone's recommendation list, over and over, no matter how different the listeners were, while other perfectly good songs never appeared at all.
Due to a problem of measuring distances in high dimensions, hub objects are recommended over and over again while anti-hubs are nonexistent in recommendation lists, resulting in poor reachability of the music catalogue.
This wasn't a bug in that one system. It was, researchers had already shown,
"a natural consequence of high dimensionality and as such is yet another aspect of the curse of dimensionality."

That single finding should unsettle anyone who assumes "personalized recommendations" describes what vector-based systems actually deliver. The industry's pitch is that your Amazon suggestions, your Spotify queue, your Netflix row of "Because you watched…" are built from a mathematical model of you — a point in space, surrounded by the things most like you. The uncomfortable truth buried in two decades of recommender-systems research is that the geometry itself is rigged, in a precise, provable way, against individuality. The math doesn't converge on your taste. It converges on a small set of mathematically privileged points that everyone gets shown, and it does so more aggressively the more "sophisticated" — meaning higher-dimensional — the system becomes.

From Item to Item, 1998 to Now

The idea of representing a shopper, a song, or a product as a vector isn't new dressing on an old idea; it's close to the founding idea of modern e-commerce recommendation. Amazon's Greg Linden, Jay Jacobi, and Eric Benson filed a patent in 1998 —
"That patent was filed in 1998 and issued in 2001"
— that became the basis of the item-to-item collaborative filtering described in their 2003 IEEE Internet Computing paper.
By comparing similar items rather than similar customers, item-to-item collaborative filtering scales to very large data sets and produces high-quality recommendations.
The trick was to stop representing customers as giant vectors of every product ever browsed and instead build item vectors, so that "similar to what you bought" could be computed once, offline, for the whole catalog rather than freshly for every shopper.

A decade later, the same geometric instinct got a huge assist from natural-language research.
Word2vec embeddings were introduced in 2013 by Google
, originally to place words in a space where "king" minus "man" plus "woman" lands near "queen." Recommendation engineers noticed almost immediately that a browsing session or a shopping cart looks a lot like a sentence — a sequence of items instead of words — and item2vec-style models were born, embedding products into the same kind of vector space that once held vocabulary. By the mid-2010s, YouTube's engineers had gone further still, splitting the problem into two neural "towers" — one for the user and context, one for the candidate video — trained so that a good match sits close together in a shared embedding space, a structure now standard across large-scale retrieval systems. Even that architecture carried its own cautionary tale:
relying heavily on features such as 'click-through rate' can result in recommending clickbait (i.e., videos users rarely complete), as compared to features like 'watch time' which better capture a user's engagement.
Vectors, in other words, only ever encode what you tell them to optimize for — and the industry has repeatedly discovered, after the fact, that it told them to optimize for the wrong thing.

Netflix's Own Engineers Break the Ruler

Here is the detail almost no general-audience piece on recommendation engines mentions: the very metric used to compare vectors — cosine similarity, the workhorse "how close are these two points" calculation cited in nearly every explainer — was shown in 2024 to be, in certain common setups, essentially arbitrary. A team led by Harald Steck at Netflix, working with Cornell's Nathan Kallus, examined the linear matrix-factorization models that sit underneath much of recommendation, and found that cosine similarity can yield results that depend on incidental modeling choices rather than any real notion of closeness. As one write-up of the paper put it,
"the research team, led by Harald Steck, Chaitanya Ekanadham, and Nathan Kallus, has uncovered a significant issue: in certain scenarios, cosine similarity can yield arbitrary results, potentially rendering the metric unreliable and opaque."
A company built on the premise that it could mathematically quantify "you'll like this because you liked that" published research undermining the ruler everyone uses to measure "like."

The Hub That Ate the Catalog

The cosine-similarity wobble is a crack in the foundation. The hubness problem is closer to a law of physics. As dimensions increase, distances between points stop varying much — everything drifts toward roughly the same distance from everything else — and a small number of points, usually those sitting nearer the statistical center of the whole distribution, end up as the "nearest neighbor" of an implausibly large share of all other points.
Hubs are data points which keep appearing unwontedly often as nearest neighbors of a large number of other data points… as the same similar objects are found over and over again and other objects are never recommended.
Crucially, this isn't a defect of any one dataset or company's engineering: it's
"a general problem of machine learning in high-dimensional spaces"
that shows up in music, images, and text retrieval alike. Add more features to a recommendation model to make it "understand you better" — more listening history, more browsing signals, more embedding dimensions — and, perversely, you often make the hub problem worse, not better, because distance concentration intensifies with dimensionality. The fix isn't more personalization data; more data along more axes is part of what causes the convergence in the first place.

When "Similar" Just Means "Popular"

If hubness explains why the same items keep resurfacing, a 2024 analysis explains which items win that lottery. Researchers studying embedding-based recommenders found that the dominant direction in the learned item-embedding space — the one axis that explains the most variance — lines up almost perfectly with plain old popularity.
The principal right singular vector of the matrix aligns significantly with the item popularity, with the cosine similarity consistently surpassing 0.98 across multiple representative recommendation models and datasets.
A correlation above 0.98 is not a subtle statistical tendency; it is close to saying the two quantities are the same thing wearing different names. When a system tells you an item is "similar" to what you like, in a great many production models it is substantially reporting that the item is popular, full stop — a fact laundered through enough linear algebra to look like insight about you specifically.

Living With a Machine That Can't Help Averaging

None of this means the engineering is dishonest. Teams that build these systems know about hubness and popularity collapse and have spent years building patches — local scaling, mutual proximity graphs, two-stage architectures that separate a broad "candidate generation" pass from a narrower "ranking" pass precisely so the hub effect in stage one doesn't dictate the final list, approximate-nearest-neighbor libraries like FAISS tuned to trade a little accuracy for enough speed to search billions of vectors in milliseconds. These are real, useful mitigations. None of them repeals the underlying geometry.

What they add up to is a quieter claim than the marketing suggests. A recommendation vector is not a portrait of your taste; it is a compressed, popularity-skewed, occasionally arbitrary coordinate that happens to sit near a gravitational center shared by millions of other coordinates. The industry calls the output personalization. The mathematics, when you read the actual papers instead of the product page, is closer to a very expensive way of finding the crowd you already belong to — and quietly nudging you to stay in it.

References