How the Genre Explorer Works

Methodology and editorial principles

The challenge

Mapping the relationships between musical genres is a problem with no clean solution because the data is simultaneously a tree, a network, a chronology and a geography. Genres descend from ancestors, which is a tree structure. They absorb influences from across family lines, which is a network structure. Every lineage claim is a temporal claim, so the data is a chronology. And scenes emerge from specific places and conditions, which makes it a geography. Any single projection must choose which of these dimensions to express and which to suppress. A tree cannot hold a genre with two parents. A network diagram shows everything yet explains little. A timeline loses lateral connections. A map hides the history.

The projects that have attempted this at scale, and there are a number of serious examples worth engaging with, all eventually ran into the same limit: the data has more dimensions than any one representation can carry.

The Genre Universe dataset

The Genre Dive Genre Universe is a hand-curated dataset of over 600 genres and over 1,100 relationships recorded by type. Every relationship is a researched and reasoned claim made by me, Paul Juniper, as editor; not a similarity score or a co-occurrence pattern computed by an algorithm. Algorithmic placement can scale to thousands of genres but often fails with complicated and occasionally murky claims of cultural influence or transfer of techniques.

Relationships are classified into three display types. Ancestry: one genre descends from another, carrying forward core structural features of the parent. Fusion: a genre is born from the deliberate collision of two or more parent scenes and its identity cannot be understood without reference to all of them. Influence: a lateral connection that shaped a genre's development without being part of its direct lineage.

The graph is directed (each edge points from antecedent to descendant or from influencer to influenced), sparse by network standards, and small enough to draw in full while large enough that drawing it without structure produces an unreadable hairball.

Genres are grouped into families, the membership of which is resolved through lineage with occasional editorial override, and every genre carries a region (the geography where its founding scene originated).

The approach: multiple views

Rather than searching for a single perfect projection, the Genre Explorer provides multiple (two currently but a third is coming) views over the same curated dataset.

The Map is a 2D view arranging genres spatially using a deterministic block layout algorithm. Families are rendered as tiered columns and every genre's position is calculated from its relationship data and depth. Background shading distinguishes family territories. If the map were designed by hand, I could have made it easier to read and likely much more attractive. However, as more genres are researched for new episodes and added to the dataset, the map needed to accommodate that growth without being manually redrawn each time so placement had to be algorithmic. Editorial decisions are encoded in the relationship data (i.e. which edges exist, which family a genre belongs to and what depth it sits at) but the rendering of that map happens according to a detailed set of rules.

The Dive view renders the same data in three dimensions, using a radial ancestry model where a selected genre sits at the centre and its lineage fans outward through space. Where the Map seeks to answer "where does this genre sit in the broader landscape," the Dive answers "where did this genre come from and what did it become." Both use the same underlying dataset and selecting a genre in either view highlights the same relationships.

Editorial principles

The taxonomy is opinionated and makes no claim to neutrality. Genres are included because they illuminate the story of how music moves between cultures or because I like them personally or am curious about them; not because they meet a threshold of commercial significance. Based on my own musical biases, the dataset is global with the shows and the Genre Explorer giving no particular preference to any region.

Where a genre's parentage is genuinely disputed, the relationship data records the strongest editorial reading rather than attempting to encode the dispute. I generally seek to explore any disputes to a genre's limitations during episodes but in this dataset I have to make the calls in one direction. Relationship notes explaining the reasoning behind contested edges are maintained but not published.

The Genre Explorer is a living document. Genres, relationships and family assignments change as new episodes are researched or as my understanding deepens. Consider the version of the dataset rendered here as a snapshot, not the final word.

Prior art

The Genre Explorer was built with awareness of some notable prior attempts at genre visualization.

Every Noise at Once — Glenn McDonald's Every Noise at Once (2013 onward) is the reference point for the algorithmic approach. Thousands of Spotify genres are placed on a two-dimensional scatter by audio attributes. Impressive in scale and fun to explore, the approach however has no visible reasoning to proximity and its design meant there are no drawn relationships between genres nor any historical dimensions.

Musicmap — Kwinten Crauwels' Musicmap (2016 onward) is the strongest prior attempt and the only one with a published methodology of any note. A hand-curated genealogy of approximately 230 main genres arranged in super-genre columns with time flowing downward and typed links connecting them. Its limits are the mirror of its strengths: the fixed canon tops out near 230 genres, hybrid genres are compromised by the column structure, and the scope explicitly excludes world and folk music. The territory Musicmap declines to cover is precisely the territory where Genre Dive focuses.

The Music Machinery experiments — Paul Lamere's projects (2011 to 2013) are an honest account of a skilled data visualization practitioner experimenting with algorithmic approaches. Each experiment found a different local solution and each demonstrated the same underlying problem: incremental local reveal keeps a similarity graph readable, at the cost of ever seeing the whole picture.

Ishkur's Guide to Electronic Music proved that time as an explicit axis works, that embedded audio transforms engagement and that an opinionated narrating voice is a feature rather than a bias problem.

Acknowledgements

This Explorer builds on and learns from every project cited above. Particular respect is due to Kwinten Crauwels for both his excellent work and publishing Musicmap's methodology chapter; to Glenn McDonald for demonstrating what totality looks like; and to Paul Lamere for documenting his experiments with a candour that made them genuinely useful to those of us who came afterwards. I thank each of you.