Inside Cultural Contraband: Music Discovery Through Graph Theory
Streaming algorithms are good at finding more of what we already recognize. Cultural Contraband Scout was built to find the cultural relationships that have not yet been priced into attention.
Most music-discovery systems begin with similarity: people who listened to this also listened to that; this song sounds like the songs already in your rotation; this artist shares an audience with an artist you follow.
Cultural Contraband begins with a different question:
Who is connected to whom—and what does the structure of those relationships reveal before the wider culture notices?
That question turned a weekly playlist into a graph problem.
Cultural Contraband Scout is a reusable research system I built for discovering new releases. It checks artists who have already appeared in Cultural Contraband, artists I follow on Spotify, music submitted through InAvator, albums and EPs, unfamiliar countries and scenes, and the broader networks around all of them. It then ranks the resulting candidates by an ex-ante theory of “cool”: what their cultural position looks like now, before streams, press coverage, virality, or hindsight make the answer obvious.
Its governing rule is simple:
Cultural Contraband should discover culture, not merely maintain its own canon.
The hard part was turning that principle into a repeatable method.
A playlist is not a list. It is a graph.
The obvious way to scout from an existing playlist is to treat its artists as a checklist. Did anyone already featured in Cultural Contraband release something this week? If so, add it to the pile.
That approach is useful, but it is not discovery. It gradually turns a playlist into a closed canon: the same artists recur because previous inclusion becomes the main qualification for future attention.
The Scout instead treats every historical artist as an identity seed in a larger graph.
In graph-theory notation, the music world becomes G = (V, E). The vertices are not only artists. They can also be:
• producers and production collectives;
• labels and imprints;
• groups and creative collaborators;
• cities, diasporas, and local scenes.
The edges are equally varied. A feature is one kind of edge. Repeated production work, membership in a collective, a formative label relationship, or a bridge between two regional scenes can be more informative than a single guest verse.
The historical Cultural Contraband registry currently supplies 369 source tracks and 468 normalized credited artists. Of those artists, 361 appeared only once and 107 recur. That is not merely an archive of past selections. It is a set of 468 possible entry points into a much larger cultural network.
The distinction matters because the track that introduced an artist to the playlist is evidence of identity, not a boundary around that artist’s career.
Suppose a Cultural Contraband track exposes Pyrex through credits involving Shiva and Drillionaire. A track-level crawler might stop there. But a whole-career graph should also surface Tony Effe through Pyrex’s repeated Dark Polo Gang relationship and Sick Luke through his formative and recurring production role. Neither connection needs to appear on the exact song that entered the playlist. The song tells the system which Pyrex we mean; Pyrex’s career tells it where to look next.
The same logic turns Tokyo into a path toward Kermit through their co-credit on “Prostituta,” but also toward Tradez through production work elsewhere in the catalog. Baby Gang can lead to Simba La Rue even if Simba already exists elsewhere in the historical registry. The release check is deduplicated, but the second path is preserved because multiple independent edges tell us something about network structure.
The edge needs a meaning
Graph expansion can become useless very quickly. A sufficiently connected artist has thousands of technically valid relationships, most of which carry little cultural information. Counting them all would reward metadata abundance rather than authority.
So the Scout uses a bounded breadth-first search. For each seed, it examines three scopes:
1. the exact track or tracks that appeared in Cultural Contraband;
2. current and recent catalog relationships;
3. repeated or formative relationships across the artist’s whole career.
It prioritizes one-hop relationships and follows a second hop only when the first supplies a concrete signal: recurring credits, a meaningful collective, an active label cluster, a strong co-sign, or a genuine bridge into another scene.
More importantly, every edge is typed and dated. Was the relationship a one-off feature or recurring collaboration? Was it current, formative, historical, or of unclear status? Did the artist write, produce, perform, or merely stand near somebody important? Is the source an actual credit, an official roster, or an inference somebody made from aesthetics?
This is where graph theory becomes cultural analysis. Raw degree centrality—the number of connections a node has—is not enough. The useful signal is closer to quality-weighted, role-aware, temporal centrality.
Three intuitions matter most:
• Weighted degree: How many credible relationships exist, how strong are they, and does the artist make a substantive contribution?
• Betweenness: Does the artist connect clusters that would otherwise remain separate—languages, cities, diasporas, labels, or scenes?
• Temporal relevance: Is the edge active now, recurrent across time, or merely part of an important lineage?
A broad distributor shared by two artists is not the same as a creative relationship. A social follow is not a collaboration. A single celebrity feature is not a scene. Ten weak edges should not outweigh two culturally decisive ones.
The system therefore does not ask only whether two nodes are connected. It asks what kind of cultural information is carried by the connection.
Why ordinary recommendation systems miss this
Collaborative filtering is extremely good at recognizing consumption patterns. Audio models are good at recognizing sonic resemblance. Popularity systems are good at identifying records already accumulating attention.
Those are all useful, but they answer variations of the same question: What is statistically adjacent to something already known?
The Cultural Contraband graph is trying to answer something else: Where is cultural authority forming before audience consensus catches up?
Sound alone cannot tell you that a producer quietly anchors a scene. Co-listening alone cannot distinguish an artist who helped create a movement from one who arrived after its codes became globally legible. Popularity alone cannot tell you whether a small artist is genuinely under-recognized or simply obscure.
This is also why the Scout forces an outside-network pass. A graph built only from past Cultural Contraband artists will eventually become very good at reproducing Cultural Contraband. It has to deliberately search countries, local-language sources, labels, and scenes that are not already dominant in the graph. Otherwise, “network intelligence” becomes a sophisticated name for path dependence.
The known network creates depth. The outside pass preserves surprise.
“Cool” is not a synonym for small
Once the discovery graph was working, a more difficult problem appeared: how should the Scout decide what deserves to be heard first?
The working definition became:
Coolness ≈ scene authority + discovery surplus + cultural specificity + low need for external validation − obvious trend-chasing.
Get Jacob N Borden’s stories in your inbox
Join Medium for free to get updates from this writer.
Each factor is scored from 0 to 10, producing a 0–100 score:
Coolness = clamp(0, 100, 2.5 × (A + D + S + V − T))
Here, A is scene authority, D is discovery surplus, S is cultural specificity, V is validation independence, and T is the trend-chasing penalty.
The arithmetic is there to enforce consistency, not to manufacture objectivity. Every score also carries an evidence note and a confidence level. When the evidence is thin, the honest answer is a range, a tie, or a wildcard—not fake precision.
The five factors solve different problems.
Scene authority asks whether the artist possesses real standing, agency, or embeddedness within a cultural setting. Repeated work with respected peers may be evidence. So can a meaningful producer role, organic local uptake, or a position that connects separate clusters.
Discovery surplus asks how much useful information Cultural Contraband adds by presenting the record now. Has the broader audience already priced in the artist’s importance, or is there a gap between the network signal and public attention?
Cultural specificity rewards work that comes from a legible place, language, lineage, community, or hybrid scene. It does not reward foreignness by itself. Geography without evidence is tourism, not specificity.
Validation independence asks whether the cultural case survives after removing chart positions, celebrity approval, prestigious press, and later success.
Trend-chasing penalizes late adoption of an already-legible sound or image when the artist contributes little mutation, agency, or local logic. It does not punish an artist for participating in a trend they helped create.
The model makes two points that recommendation culture often collapses.
First, popularity is neither a reward nor a penalty. It is evidence about visibility and consensus.
Second, obscurity is not coolness. Low visibility becomes interesting only when other evidence suggests that something culturally important remains under-recognized.
Network leverage or borrowed prestige?
This is the most delicate part of the model.
An artist with a small audience and much larger collaborators may be sitting on unpriced cultural capital. Or the artist may simply be borrowing the glow of one famous relationship. The graph has to distinguish embedded authority from dependency.
Conceptually, the Scout looks at network evidence like this:
Network evidence ≈ Σ(edge strength × neighbor authority × candidate contribution)
That evidence becomes especially interesting when it is paired with a large visibility gap: the artist is far less recognized than the credible network around them would predict.
Daxter is a useful example. If Daxter has substantially fewer listeners than Il Ghost and Stresi while holding verified, substantive relationships to them, that gap matters. Multiple meaningful edges may indicate that Daxter is more embedded than his broad visibility suggests. But the conclusion still depends on role, recurrence, and independent evidence. Listener count alone proves nothing.
The model uses a few counterfactuals to keep itself honest:
• Remove the audience numbers. Whose cultural position still makes sense?
• Remove the largest collaborator. Whose authority survives?
• Compare the strength of the network with current visibility. Whose position appears least priced into attention?
• Remove later success. Would the discovery have been defensible at the time?
• Remove familiarity to the curator. Is the artist cool, or merely recognizable?
These tests explain why my calibration can say Solange is cooler than Beyoncé without pretending Beyoncé lacks authority. Solange’s independent cultural position, specificity, and lower need for mass validation can outweigh Beyoncé’s scale.
They also explain why Memphis Bleek is not cooler than Jay-Z simply because he is less famous. If removing Jay-Z collapses the case, obscurity plus proximity is borrowed prestige, not autonomous cool.
Likewise, Simba La Rue can rank above Baby Gang even though Baby Gang may possess greater reach or broader authority. Coolness is not a fame leaderboard inverted. It is a judgment about the combination of authority, specificity, discovery surplus, and independence at a particular moment.
The graph produces a listening queue, not a verdict
The full scouting funnel runs in a fixed order:
Past Cultural Contraband seeds → Spotify follows → whole-career network expansion → outside-network releases → InAvator submissions → album and EP mining → coolness ranking → listening → sequencing
Ranking happens only after the discovery passes are complete. That prevents the first familiar names or heavily promoted singles from anchoring the entire search.
The resulting records are divided into three audition bands: listen first, strong possibilities, and wildcards. None of them is called a selection before listening.
That boundary is essential. Graph theory can tell me why a record deserves attention. It cannot tell me whether the record actually works inside the edition. Metadata can reveal a cultural bridge, a surprising producer, or an underpriced network position. It cannot replace the ear.
The Scout automates research discipline, not taste. Its job is to make the listening pool wider, deeper, more traceable, and less vulnerable to whatever an algorithm or publicity campaign happened to place at the top of the screen that morning.
The same principle applies when judging whether the system worked. PIXY before “LEGACY” broke or 3RMZZ at roughly 40 listeners can help calibrate the method because they show the value of detecting cultural signals before consensus. But their later trajectories cannot be retroactively inserted into the original score. That would be hindsight bias.
A serious discovery model also has to remember the early finds that never broke out. Otherwise, it becomes a machine for retelling successful predictions and forgetting the misses.
Music discovery as cultural price discovery
The economic analogy is useful: attention behaves like a market, but not an efficient one. Information moves unevenly across languages, cities, platforms, and scenes. A producer may be locally central long before becoming globally visible. An artist can accumulate strong cultural relationships without those signals reaching a broad audience. A deep cut may reveal more directional movement than the single carrying the marketing budget.
Cultural Contraband Scout looks for those mispricings.
It does not assume that the smallest artist is the best discovery, that every foreign scene is inherently interesting, or that adjacency to a star creates authority. It looks for a rarer combination: credible embeddedness, substantive contribution, cultural specificity, independent logic, and a level of recognition that has not yet caught up.
That is what “cool” means inside Cultural Contraband. Not popularity. Not obscurity. Not personal taste disguised as mathematics.
It is the cultural information contained in the gap between where an artist sits in the graph and how clearly the wider world can see them.
A streaming service asks, “What are you most likely to play next?”
Cultural Contraband asks, “What is happening in culture that you do not know to ask for yet?”
