Is Hulk an accident?

The same four tests on Wikipedia's Category: Marvel Comics supervillains — 370 nodes where the heroes Hulk, Deadpool and Doctor Doom loom largest, nearly 40% of citations run both ways, and 43 villains nobody links to at all.

This is the network behind the crawl in data/Marvel_Comics_supervillains: the 370 pages reachable in Wikipedia's Category: Marvel Comics supervillains (redirects resolved; crawled 2026-09-13), with 1,136 directed edges between them — an edge A→B when B's article appears on A's. Mean in-degree is 3.07, barely half the hero network's 5.89, and the structure is frayed at the edges: 46 weakly connected components, 43 characters nobody links to in either direction, a giant component of 323, and two genuine two-villain islands — Humus Sapien with Ogre, and Tetu with Zenzi.

The hub race has a plot twist: the most-cited "supervillains" are mostly heroes. Hulk tops the count at 58 in-links, Norman Osborn follows at 50, then Deadpool (25), Thanos (23), and Black Cat and Venom tied at 22 — category leakage, not a labeling error, since villains' articles genuinely cite their heroic nemeses. One genuine stowaway does hide in the node list though: David 8, the android from the Alien franchise, present and fully isolated — the same kind of crawl artifact as Street Fighter's Tifa Lockhart. Is a hub of 58 just what scattering 1,136 links across 370 nodes produces, or does it take something more deliberate? Same order as always: chance, degree-preserving chance, growth, and the individual level.

Erdős–Rényi first: the same 370 nodes and 1,136 directed edges, placed by independent coin-flips. Across 300 offline draws, the biggest hub's in-degree ranged from 7 to 15 (mean 9.36, standard deviation 1.05). Hulk's 58 is above every single draw — about 46 standard deviations above the random mean, and nearly 4× the most extreme draw of 300. Norman Osborn's 50 clears the same ceiling just as decisively. No ambiguity here: chance doesn't build hubs like these. Draw your own random network below and watch its biggest hub land.

The tighter null: keep everyone's exact in- and out-degree fixed, scramble only who links to whom — repeatedly take a chain a→b, b→c, c→d and rewire it to a→c, c→b, b→d. Two statistics:

• Clustering coefficient — the average local clustering coefficient on the undirected projection (an edge exists between two characters if either one's article links to the other's), averaged over all 370 nodes; degree-<2 nodes contribute 0. Real value: 0.2064.
• Reciprocity — mutual directed edges ÷ total directed edges = 452 ÷ 1,136 = 0.3979.

Across 300 offline shuffles (3,000 swaps each): clustering's null mean was 0.0763 (sd 0.0069) — the real value sits about 19 standard deviations above it, beyond all 300 shuffles (empirical p ≈ 0). Reciprocity's null mean was 0.0481 (sd 0.0079) — the real value sits about 44 standard deviations above that, again beyond every shuffle. Degree alone doesn't explain this network's triangle-heavy structure or its unusually mutual citation habits — villain articles name each other back at a rate random wiring can't touch.

The chart below starts pre-populated with shuffles, so the null distribution is visible immediately — pick a statistic, or run more yourself.

Preferential attachment: grow to 370 nodes one at a time, and give each newcomer m outgoing links to existing characters — chosen with probability proportional to how many in-links they already have. The classic rich-get-richer mechanism, in its directed citation-network form (the Price model), which fits our directed network better than the usual undirected Barabási–Albert setup. Out-degree is exactly m for every non-seed node by construction; in-degree accumulates by preference.

This is a generative hypothesis tested against the data — not a claim about how the Wikipedia article network actually got written.

At m=3 (3×370 = 1,110 edges, close to the real 1,136), 50 offline trials all produced a bigger hub than Hulk: max in-degree 86–175 (mean 120) against the real 58 — preferential attachment doesn't just reproduce a big hub here, it more than doubles it. Even at m=2 (mean 88, range 55–113) every trial overshoots; at m=4 the gap widens further (mean 137). Lowering m brings the hub down but pulls the edge count further from 1,136 — no single m matches both density and hub size. And at every m, every trial: one connected component, zero isolates — versus the real network's 43 isolates and two two-villain islands. A model where everyone who joins immediately attaches can't produce a villain nobody ever mentions.

Move the slider and regenerate to see for yourself.

A villain universe born all at once

Preferential attachment depends on history: characters arrive one at a time, and each newcomer is more likely to connect to whoever's already popular, so an early lead can snowball into a hub like Hulk. But what if the villain universe was born all at once?

If all 370 characters existed simultaneously, there'd be no established popularity for a newcomer to follow — everyone would start with the same shot at new connections. That's the counterfactual below: a second model, same 370 characters and roughly the same edge budget, that links nodes uniformly at random instead of preferring the well-connected. It's not a claim about how Marvel's Wikipedia network actually came to exist — it's a way of asking how much of the hub structure could come from history alone.

Run each model a few times at the same m. Big Bang spreads connections fairly evenly — most nodes end up a similar, modest size, and the largest hub barely varies run to run. Sequential growth, at that same m, reliably produces one or two stand-out characters, because whoever gets ahead early keeps attracting a disproportionate share of new links. That's the mechanism in miniature: Hulk's 58 real in-links aren't just abstract "popularity" — in a growing network, being popular early is itself an advantage that compounds. In a universe born all at once, there's no "early" for that advantage to attach to.

None of this shows preferential attachment is what actually produced the real Wikipedia network — the CCDF comparison above already found PA, at the edge-matched m, reliably overshoots Hulk's real in-degree. What it does show is why network history can matter: a heavy-tailed degree distribution is consistent with a growth-plus-preference story, but consistent isn't proof — this counterfactual just isolates what history specifically contributes.

Definitions: on the undirected projection (same one used for clustering above), a character's degree is their number of distinct neighbors, and their average neighbor degree is the mean degree of those neighbors. Network-wide, ⟨k⟩ = 4.919, while the edge-weighted mean neighbor degree ⟨k²⟩⁄⟨k⟩ = 13.18 — nearly triple. That gap is the friendship paradox. It's worth stating plainly that ⟨k²⟩⁄⟨k⟩ is a pure function of the degree sequence, so it's mathematically guaranteed to survive a degree-preserving shuffle — not something this page needs to test.

What isn't guaranteed is what it looks like character by character. Of the 327 connected characters, 274 (83.8%) have neighbors who are on average more connected than they are; 44 (13.5%) out-rank their average neighbor, and 9 are exactly tied. Fifteen characters are never out-popularitied by any neighbor at all — the local degree champions, from Hulk at the top of the giant component to Ogre and Humus Sapien inside their own two-villain island, each merely the biggest fish in a pond of one relationship.

Search for a villain below to see their own numbers.

We plotted the in-degree CCDF P(K ≥ k) on the frozen snapshot (370 characters, 1,136 directed links) against two nulls: an ER random graph, whose in-degree is exactly Poisson with λ = 1,136/370 ≈ 3.07, and a 200-seed ensemble of directed Price graphs (N = 370, m = 3 — edge-matched), with the ensemble maximum marked. A discrete power-law fit (Clauset–Shalizi–Newman MLE, via the powerlaw package) then summarizes the tail: exponent α̂ with its standard error, and a self-selected cutoff k̂_min.

• Mean in-degree ⟨k⟩3.07
• Max in-degree58 · Hulk
• Fitted exponent α̂2.43 ± 0.15
• Fitted tailk ≥ 4 · 94 of 267

Not random. Under ER with λ = 3.07, a character with in-degree 12 or more should essentially never exist — expected count 0.033 across all 370 nodes, real count 20. A degree of 58 is off the Poisson scale by 1.3 × 10⁻⁴⁹ expected occurrences in a network of this size: the gray curve exits the chart by k ≈ 12, while the real network still has 7 characters at k ≥ 20.

Lighter than Price. The fitted exponent (α̂ = 2.43) sits well below Price/BA's theoretical 3.0 (smaller exponent = heavier tail), but the ensemble comparison points the other way at the extreme: every one of the 200 Price seeds produced a hub above Hulk's 58 (ensemble max 164, mean 118). Forced to choose between random and Price: Price, easily — it just overstates this network's skew at the very top, the same overshoot the growth section found.

Neither, strictly. The power law describes the tail, not the whole network: only 94 of the 267 cited characters sit in the fitted region, 103 characters were never cited at all, and a network this small carries wide uncertainty on α̂. "Scale-free-looking tail" is as far as this data goes.

What we can't conclude (for this question):

• The mechanism. A heavy tail doesn't prove preferential attachment — many processes produce one, and the frozen snapshot has no timestamps to test Price's growth story at all.
• The exact exponent. α̂ comes from just 94 tail points; everything below k = 4, including the 103 never-cited characters, is outside the fit.
• A like-for-like Price test. The comparison is directed-in-degree against a directed model, but the shape verdict is still visual, not a fitted-model likelihood comparison.

Pulling the four sections together:

• Independent random edges (ER) are rejected outright. 300 draws topped out at max in-degree 15 — nowhere near Hulk's 58 (z ≈ 46).
• Degree alone doesn't explain the clustering or reciprocity either. Both sit far outside the shuffle's null distribution (z ≈ 19 and z ≈ 44) — real relational structure (teams, mutual nemeses, shared storylines) beyond what link count alone predicts. The one thing that stayed fixed under the shuffle — the giant-component/islands split — is invariant by construction, not a discovery about the network.
• Preferential attachment overshoots harder here than anywhere else so far. Every trial at every tested m more than doubled Hulk's real in-degree, and the 200-seed ensemble never once produced a hub as small as 58. The model structurally can't produce isolated characters or disconnected islands either: everyone who joins, connects — versus this network's 43 isolates and two two-villain islands.
• The friendship paradox holds, strongly, for two different reasons. Network-wide, ⟨k²⟩⁄⟨k⟩ (13.18) nearly triples ⟨k⟩ (4.919) — guaranteed by the degree sequence alone. Individually, 83.8% of connected characters personally experience it — but who plays "the popular friend" for whom depends on the network's actual wiring, not just link counts.

So: no, Hulk isn't an accident of scale or density. Something like rich-get-richer growth fits the real hub structure far better than any random baseline — but that's a hypothesis this data is consistent with, not something these four tests prove happened, and even its best-fitting version overshoots the real hub while structurally ruling out the 43 characters nobody bothered to link to. The villains' verdict rhymes with the heroes': heavy tail, real structure, model in the right neighborhood — and one stowaway android from a different franchise entirely.