When I first began building WhatNext., the simplest way to model someone's movie taste seemed almost too neat: take the mathematical fingerprint for every movie a person likes, calculate the average across all of them, and use that single point as their taste profile.
To suggest what to watch next, the recommendation engine would just search the catalogue for films whose tone and style sat closest to that average.
If your movie watching strictly orbits a single aesthetic (say, brooding 1970s political thrillers), that single-point average works reasonably well. But almost nobody actually watches movies that way. Most people have tastes split across distinct aesthetic islands: bleak, atmospheric sci-fi on Friday night, and high-energy comedies or family adventures on Sunday afternoon.
The moment you blend those contrasting moods into a single average, the math quietly breaks down.
The Trap of the Mathematical Middle
In a recommendation engine, films are represented by coordinates that capture tone, pacing, theme, and genre. Movies with similar artistic textures cluster close together in that space.
When you take the average of two opposite aesthetic worlds, it does not land on a film that captures both. It lands directly in what I call the semantic dead zone: the empty space between them.
Consider what happens if you love heavy, meditative science fiction like Dune and surreal, stylized dramas like Poor Things, but you also genuinely enjoy lively animated comedies like Kung Fu Panda.
When an algorithm averages those coordinates together, the dark, slow-burn qualities of Dune pull against the bright, kinetic rhythm of Kung Fu Panda. They cancel each other out. The resulting average describes a lukewarm, mid-energy drama: something neither too fast nor too slow, neither dark nor light. It represents nothing you actually love.
This phenomenon is called vector cancellation. The films sitting at the outer boundaries of your taste, the very titles that define your most passionate viewing moods, get buried because they sit too far from that diluted center point.
Catching the Flaw on a Real Watch History
Early in WhatNext.'s development, I wanted to find out whether vector cancellation was an abstract theoretical quirk or a genuine defect hurting real recommendations.
To test it, I analyzed an export of my own watch history, looking closely at how my positively rated films scored against a single blended profile versus distinct taste hubs.
The defect was immediately obvious.
The titles sitting furthest from my blended average were the very ones that defined my lighter moods. Favorites like Kung Fu Panda, Moana, and Lilo & Stitch scored near the bottom of my own library. Because my watch history leaned heavily toward sprawling sci-fi like Dune and monster epics like Godzilla, the mathematical average was dragged so far into serious, dark territory that lighter films were virtually invisible to the recommendation engine.
Even though I had rated those films positively, the algorithm quietly behaved as though they didn't belong to me at all.
Why Clusters Form on Tone Rather Than Format
When I first looked at those results, it was tempting to assume the algorithm had simply failed to understand animation. But looking deeper into how films relate to one another revealed something more subtle.
The coordinates that describe a film encode narrative texture, emotional tone, and energy, not whether a film was shot on 35mm or rendered on a computer. The grouping didn't separate films because of their production medium; it grouped them by their shared warmth, pacing, and comedic timing.
Within that lighter cluster, Kung Fu Panda and Madagascar sat closer to each other than to Moana or Lilo & Stitch, united by their zany, kinetic rhythm. Live-action family comedies with similar energy levels could easily share that exact same space.
Recognizing distinct aesthetic hubs lets each side of a person's taste breathe on its own terms, without lighter favorites having to fight darker epics for mathematical dominance.
Letting Each Mood Advocate for Itself
Today, WhatNext. maintains distinct taste hubs directly on your device rather than forcing your watch history into a single average.
When the recommendation engine searches the catalogue, a film is scored against all of your taste hubs and judged by whichever mood it matches best. If a new sci-fi film is a home run for your Dune hub, it gets recommended immediately, even if it has zero connection to your Kung Fu Panda hub. Each of your viewing moods gets to advocate for what it loves.
Human taste is not a single point on a map. It is an archipelago of moods, genres, and memories. By honoring those distinct islands rather than forcing them into an artificial compromise, the math finally matches the way people actually watch movies.