Well, Someone Has to Write About You First

New artists, obscure artists, non-Anglophone artists are all underrepresented in the Stell-R universe, simply because fewer critics have covered them. Further, the supply of real long-form music criticism keeps shrinking as outlets close. We can't really force this gap to close.

But can we make it better? Notice how audio always exists. Every artist has tracks, and tracks have measurable acoustic features, regardless of whether anyone ever reviewed them. So: can we predict critical adjacency from sound alone, for artists no critic has ever connected?

Testing It Properly

Hypothesis: we can decently recover critical similarity with sonic similarity. To do this, instead of reducing each artist to one "average sound," I compared their full distribution of tracks across 80 audio features (timbre, rhythm, harmony, and more), using optimal transport (i.e. the math used to measure the cheapest way to reshape one pile of sand into another). Then I use those 80 pairwise distance as predictos to impute the presence of a critical edge in unlabeled artist pairs.

>

Then I trained and tested this on an artist-disjoint split (so no artist in the test set was ever seen during training).

And it performance barely moves across how well-covered an artist already is, meaning the signal survives where the graph is sparse. It's also not secretly just a genre detector since we show it discriminates fine even within a single genre.

And the Leftovers? Food for Thought!

Now, not true (at all) that whenever a critic links two artists it has to do with sound (like, at all). A lot of it is persona, scene, lineage, cultural narrative...none of that necessarily lives in the audio itself. So this is merely recovering the sonic core of critical judgment, while leaving the rest (the sociology-of music part, if you will) as its own open question (and an extremely interesting one!), rather than a residual error to explain away or ignore.

This was presented as a poster at USRW 2026 (Unified Search and Recommendation Workshop, co-located with ACM RecSys 2026, Minneapolis). Full paper, poster, and BibTeX are up on the research page.

↑ Show less

First of all: What's in a Stell-R Connection?

To recap (more about this on our last post), every connection (edge) between two artists (nodes) in the Stell-R graph means that at some point they were linked by music experts in their writing: one artist's name showed up (non-randomly) in the other's musical reviews. We take that as a powerful measure of artistic influence. In addition, for each of such edges we store two attributes:

- Number of mentions—how many documented sources (reviews, interviews, liner notes, music history writing) cite that one artist influenced the other. So a connection with 40 mentions represents a lot more critical consensus than one with 2.

- Sonic similarity—a score from 0 to 1 for how similar the two artists actually sound, based on comparing the distribution of their songs' audio features (tempo, energy, danceability, and so on). 1 means close to acoustically identical, 0 means nothing alike.

Fun fact: these two numbers frequently disagree. Two artists can also just happen to sound alike with zero documented lineage between them. But each (albeit independently) carries information that is extremely valuable for our algorithms.

Now, at this point it's only natural to wonder: well what's inside this graph? what's its structure? how are A and B connected? how is everyone connected?? Well, that's precisely what Stell-R is about.

Quick Path: Fewest Hops, the "Efficient" Road

"Quick path" is based on a simple yet monumentally important graph algorithm: the shortest path. Our implementation (Djktra's) runs a breadth-first search, in the spirit of a "six degrees of separation" game. It treats every connection as exactly the same and asks one question: what's the fewest number of hops from artist A to artist B? Number of mentions don't factor in at all, so a connection two people ever wrote about counts exactly the same as one forty people have written about.

Sonic similarity does sneak in, but only as a tiebreaker: among several equally-short routes, Quick path prefers whichever one passes through the most acoustically similar artists. It's a tiebreaker, not the main criterion. The main criterion is purely "fewest stops."

Explore Deeply: The Most Agreed-Upon Skeleton

"Explore Deeply" also looks for a "best path " between your 2 artists but considers "best" as something quite different from "short". The idea: for every one of the ~25,000 artists in our Stell-R network, keep only their single most heavily-documented connection, and discard the rest. Rinse, repeat...what's left? The smallest possible non-circular skeleton embedded in the whole graph that still connects the entire network! Graph theory calls this a minimum spanning tree.

Now, once we have uncovered that (beautiful!) skeleton, it turns out there's exactly one route between any two artists (including the ones you picked) on it. That route is Stell-R's "deep" path. Most likely the shortest one, not necessarily the most sonically consistent...but definitely the one built from the strongest documented consensus at every single step.

Cool? Let's see them both in action!

Same Two Artists, Two Different Lenses

Metallica → Beyonce

Both are "correct" by their own rules. Quick path found a real 2-hop bridge and stopped, uninterested in how well-cited that bridge is. Explore Deeply doesn't care that a shorter route exists—it's tracing metal into prog, into industrial, into ambient, into glam, into pop and hip-hop, one heavily-documented link at a time.

Fleetwood Mac → Tame Impala

Quick path's answer isn't wrong—Beach House really is a documented bridge between the two, and it's a short one. But it's a single data point. Explore Deeply's 8-hop route reflects an entire lineage of critical consensus: folk-rock introspection into baroque pop into psych, with Brian Wilson landing one hop from the destination because that connection is exceptionally well-documented, not because it happens to be nearby.

Neither One Is "More Correct"

They're answering different questions. If you want the shortest real bridge between two artists, Quick path gives you that, and a slight nudge toward the most sonically-consistent option when there's a tie. If you want the route that best reflects how critics and historians actually connect the dots—regardless of how many stops that takes—that's Explore Deeply. Same graph, same edges, two different things being optimized.

Ok. Now that you know, try them out!

Run the same pair through both on the path finder and see how differently they read the same data.

↑ Show less

What I Discovered

Spoiler Alert: Indeed, mindblowing. And it all made a lot of sense

What if I told you that Taylor Swift emerges almost directly from Joni Mitchell, Faith Hill, and The National? Or that David Bowie is the ultimate musical bridge, connecting 442 different artists across completely different genres?

That's the type of thing I discovered when I analyzed 21,243 artists and their influence connections. Instead of relying on streaming data or algorithmic guessing, I mapped the actual creative relationships between artists—and what I found was extraordinary.

Organic Musical Communities Emerge

My analysis revealed that artists in the Stell-R network organized themselves into 12 distinct communities—based purely on their review-based connections, and that that those communities made perfect musicological sense!

The Ultimate Musical Bridges

Some artists don't just influence their genre—they connect entire musical universes. These are the true innovators who shaped everything that came after:

Discoveries That Will Blow Your Mind

My network analysis revealed connections that would be impossible to find through traditional genre classification:

🎵 The Classical-Electronic Bridge: Philip Glass has 184 external connections, linking classical minimalism directly to ambient electronic music and film scores—making him the most connected classical composer in our network.

🎵 The Unexpected Pop Connection: Depeche Mode connects the pop universe to dark electronic territories with 264 external connections—more than Taylor Swift's 167 connections, revealing how electronic innovation drives pop evolution.

🎵 The Hip-Hop Experimentalists: James Blake bridges Kanye's hip-hop universe to Brian Eno's electronic experimentalism with 167 connections, while Flying Lotus adds another 126 connections between hip-hop and electronic experimentation.

🎵 The Folk-Everything Bridge: Bob Dylan doesn't just influence folk—his 287 external connections touch rock, country, indie, and even electronic music, while maintaining the highest influence score (0.155) in his entire universe.

🎵 The Beatles' Surprising Position: The Beatles appear in James Brown's soul/funk universe with 223 external connections, showing how their influence extends far beyond rock into the foundations of rhythm and groove.

Try It Yourself: Explore These Hidden Connections

Want to see these connections in action? Here are some mind-bending paths you can explore on Stell-R:

Radiohead → Taylor Swift See how experimental rock connects to pop through unexpected bridges

Miles Davis → Kanye West Discover how jazz innovation flows into hip-hop creativity

Johnny Cash → Aphex Twin The most unlikely connection you never saw coming

Explore David Bowie's Complete Network See why Bowie is the ultimate musical connector

Why All Of This Matters

Most music recommendation systems are built on listening data (e.g. what people stream together). But influence isn't about similarity, really, but about creative pathways.

When Black Sabbath influenced Mastodon, when Brian Eno shaped Radiohead, when Miles Davis inspired Flying Lotus: basically, creative genealogies.

That's why your usual algo might recommend another (perhaps super derivative) indie rock band when you like Arctic Monkeys, but Stell-R shows you that Arctic Monkeys connects to Lou Reed, to The Velvet Underground, to Andy Warhol's Factory, to an entire universe of art and creativity you never knew existed.

The Science Behind

My network analysis uses graph theoretical algorithms—breadth-first search, shortest path algorithms, and maximum flow analysis—applied to a graph of genuine artistic influence rather than streaming behavior.

I identified communities using clustering algorithms that revealed these 12 musical universes naturally, without forcing artificial genre boundaries. The result? A map of music that actually reflects how creativity flows between artists.

↑ Show less