Atlas: a map of movie taste

The first version of Suggestream worked like a walk through TMDB's recommendation graph: swipe, score neighbors, rinse and repeat. That is a great way to follow a trail from one title to the next, but it never quite answered a bigger question. What does the whole catalog look like when you step back? Where do genres bleed into each other, and which films sit at the crossroads of different kinds of similarity?

Atlas is the answer we built for that. Instead of only chaining local suggestions, we place movies and TV shows in a shared space you can explore as a map. The idea is simple to use and a little more ambitious under the hood: first learn several notions of "this feels related," then project that knowledge onto a readable landscape.

Similarity is not one thing

When someone says two films are similar, they rarely mean the same thing. Sometimes they share an audience: people who loved one also loved the other. Sometimes they sit on the same browsing path: explorers hop between them in a single session. Sometimes the kinship is in the material itself: genres, themes, and keywords that describe what the title is about.

Rather than force those ideas into a single score, Atlas keeps several embeddings in memory: vector portraits of each title, each trained for a different kind of closeness:

• Taste: titles that tend to be liked by people with overlapping preferences.
• Journey: titles that often appear together in real exploration sessions, even when the path is curious rather than pure fandom.
• Content: titles that look alike on paper: genres and thematic keywords from the catalog metadata.

Each embedding answers its own question. Taste captures preference. Journey captures how people actually wander. Content keeps obscure titles connected even when few people have swiped on them yet. Films and series live in the same spaces, so a neighborhood can mix both without special casing.

One blended memory

For recommendations and neighbor lookups, those portraits are blended into a single mixed embedding. Think of it as a compromise between gut feeling, browsing habits, and catalog DNA: each title gets a fused position that borrows from whichever signals are available. A well-known hit may lean on taste and journey; a quieter film can still find its place through content alone.

Closeness in that mixed space is what powers "what's nearby?": not a hand-tuned list of TMDB suggestions, but a comparison across the whole learned landscape. That is the memory Atlas consults when you ask for neighbors.

Then a map for the eyes

A high-dimensional similarity space is useful for math and terrible for humans. So we run a separate visualization embedding: a careful flattening of the mixed space onto two dimensions. That projection is how Atlas draws the scatter of titles you can pan and zoom.

This distinction matters. The map is a reading aid, a way to see clusters, borders, and outliers. Recommendations still rely on the richer similarity model behind the scenes. Two titles that feel related in spirit may sit a little apart on the map if the projection had to stretch the space to make room for everything else. The map shows the territory; the embeddings remember the distances.

On top of that layout, Atlas groups compact regions and labels them with distinctive themes when the data supports it. The colors follow genre as a quick visual cue, while the shape of the map comes from how titles relate across all three lenses of similarity.

Why bother?

Swiper and Explorer are still great for refining a mood one choice at a time. Atlas is for the opposite impulse: orientation. You can spot a territory, drift toward its edges, or jump to a title and inspect what sits around it, without pretending that "similar" only means "TMDB also recommended this."

The reflection behind the feature is the same one that started Suggestream: recommendation should feel like discovery, not like being trapped in a popularity loop. Multiple embeddings give us several honest definitions of relatedness. Mixing them gives the product a single memory. Projecting that memory into a map gives you something you can actually explore.

If you want to see it rather than read about it, open Atlas and wander.