Game Taxonomy

Game Taxonomy is Hushcrasher's research on the classification of video games. We learn how games group together from all the data available on each game (free-form text, images and structured data) with unsupervised and semi-supervised machine learning, instead of sorting them into predefined categories.

Today, the industry classifies games with store metadata: genres, tags and features. Most market benchmarks rely on these shared, predefined categories alone. This raises three problems.

  • Coverage. Store metadata describes genre and features only. Most of what distinguishes games has no field and no shared classification: production scope, art direction, narrative theme, the niche a game addresses, its player audience.
  • Source. Store categories are chosen by the game's developer or publisher, so they are subject to human error, differences in interpretation and marketing intent. They reflect how a game is marketed, or how its developer sees it, which can differ from how players see it. On Steam, players can add tags, but the developer sets the initial tags and their order, and can remove the tags players add.
  • Form. A predefined list has five limits:
    • It misses what is not on the list. A new niche, sub-genre, narrative theme or visual style stays invisible until someone creates a label for it.
    • Its boundaries are conventions. Classifications that go beyond store metadata (such as indie, AA and AAA) often rest on thresholds chosen by their author rather than estimated from data. These thresholds vary from one source to another and do not follow the market (an AA production in 2026 is not what it was in 2018).
    • It cannot be followed over time. When a store adds or removes a tag, the label changes but not the games: a genre can exist before its tag and after it.
    • Some of what defines a game cannot be predefined at all. For instance, classification of games by their player personas must first be learned by clustering players.
    • It puts each game in only one category whereas some dimensions are a matter of degree. For instance, a game's audience is better described by a distribution over several personas than by a single one.

Learning categories from data addresses these three problems. Any axis for which data exists can be classified. Declared labels, such as tags, become one input among many rather than the classification itself. Boundaries and memberships are estimated from the data and re-estimated over time, so that new groups can be detected as they form.

Methods

The project seeks two objects: groups of games along each axis of classification, and a distance between any two games.

Much of what describes a game is unstructured: store text, capsule art and screenshots, music. Embedding models can turn these into learnable representations, to be combined with structured data such as tags, credits, or price. Using clustering, groups can be defined on these representations, without supervision or anchored on a few reference games.

Closeness in a space that combines text and image embeddings with encoded tags, and other numerical features is not similarity by construction: it depends on how these heterogeneous features are scaled and weighted against each other, and unsupervised clustering gives no criterion for that choice. Player affinity, the overlap between the audiences of two games, provides an observed measure of which games are close. The open question is how to learn a metric on the representations whose rankings agree with these observed ones.

Papers

Applications

  • Classifying games by genre, niche, production scope, art direction, narrative, player persona, and more.
  • Finding the games most comparable to a title (beyond just store tags and features).
  • Spotting and defining a subject (such as a niche or a narrative) by the games that belong to it.
  • Tracking the evolution of clusters over time. For instance, how each genre and sub-genre cluster grows, splits or declines.

Used in

Press coverage

Polygon, Aftermath, GameDiscoverCo, Waxy.org, Critical Distance, GameLook, V2.fi, New Game Plus, Durée de vie, AppTrigger, Youxichaguan.

We license our datasets and model outputs for non-commercial academic research. Reach out →