Preferential Attachment · The Atlas of Emergence

Watch the network grow. Each new node attaches with m edges; under Preferential attachment it picks targets in proportion to how many connections they already have, so early popular nodes balloon into bright amber hubs while latecomers stay small teal leaves. Switch attachment to Random and the hubs vanish — everyone ends up roughly equal. Drag a node to reposition it.

What you're seeing

A network being built in front of you, one node at a time. Every newcomer arrives with a fixed number of links, m, and must choose which existing nodes to connect to. Under preferential attachment it does not choose at random: it is biased toward nodes that are already popular. A node with many connections is therefore more likely to gain still more — a feedback loop sometimes called cumulative advantage or simply "the rich get richer." The result is the lopsided structure on screen: a few hubs with enormous degree, surrounded by a vast majority of leaf nodes with just a link or two. The force-directed layout (springs along edges, repulsion between nodes) is only there to make the structure legible; it is not part of the model.

The rule

Two ingredients, repeated forever:

Growth. Start from a small seed of connected nodes. At each step add one new node.
Preferential attachment. The new node forms m edges, and the probability of linking to an existing node i is proportional to i's current degree k_i: Π(i) = k_i / Σ_j k_j.

That is the whole specification. A clean way to implement the bias is the classic O(1) trick used here: keep a list holding every edge endpoint — a node appears in it once for each unit of its degree — and pick a target by sampling that list uniformly. High-degree nodes occupy more of the list, so they are picked more often, automatically and in exact proportion to degree. The contrast control flips this off: under Random attachment, targets are chosen uniformly over all nodes regardless of degree. Growth still happens, but the bias is gone — and so are the hubs. Degrees stay bunched near the mean with a narrow, light-tailed (exponential) spread instead of a power law.

Why it matters

This is a mechanism for the scale-free structure famously reported in many real networks: a degree distribution that follows a power law, P(k) ∼ k^−γ (the Barabási–Albert model gives an exponent γ ≈ 3), with no characteristic scale and a fat tail of a few dominant hubs. Crucially the hubs are not designed in — no node is special at birth. They emerge from the simple cumulative-advantage bias applied to a growing system. The same structure has a sharp practical consequence: scale-free networks are robust to random failure (knock out a random node and it is almost surely an unimportant leaf) yet fragile to targeted attack (remove the few hubs and the network shatters). That asymmetry shapes how we think about everything from internet resilience to vaccination strategy.

In the wild

Cumulative advantage was first described not by Barabási and Albert but by Derek de Solla Price in 1976, who used it to explain the skewed citation counts of scientific papers — a highly-cited paper attracts still more citations. The 1999 Barabási–Albert paper rediscovered and popularized the mechanism in the language of networks, with the World Wide Web's hyperlink structure as the headline example; related claims have been made for some biological (protein-interaction, metabolic) and infrastructure networks.

Honesty check: the once-sweeping claim that scale-free networks are everywhere has been seriously challenged. In a study of nearly 1,000 real-world networks, Broido & Clauset (2019) found that strictly scale-free structure is actually rare — most networks are better described by other heavy-tailed or non-power-law forms, and clean power laws are the exception, not the rule. Preferential attachment remains an influential and genuinely useful mechanism for how hubs can arise, but it is not a universal law of networks, and this Atlas presents it as the former, not the latter.

Try this

Let a preferential network grow, then switch attachment to Random and reset: watch the hubs vanish and the graph flatten into an even mesh. This is the key teaching contrast — same growth, same m, opposite outcome.
Raise m (edges per new node). More links per newcomer means a denser core and even more dominant hubs under preferential attachment.
Watch a late arrival. Once a few hubs exist, a newly added node almost never catches up — it has too few links to attract many more. This "first-mover advantage" is baked into the rule.

For other ways networks self-organize without a designer, see the agent-and-trail Slime Mould Networks, where transport veins condense out of local reinforcement rather than explicit growth.

References

Barabási, A.-L. & Albert, R. (1999). "Emergence of scaling in random networks." Science 286(5439), 509–512.
Albert, R. & Barabási, A.-L. (2002). "Statistical mechanics of complex networks." Reviews of Modern Physics 74(1), 47–97.
Price, D. J. de Solla (1976). "A general theory of bibliometric and other cumulative advantage processes." Journal of the American Society for Information Science 27(5), 292–306.
Broido, A. D. & Clauset, A. (2019). "Scale-free networks are rare." Nature Communications 10, 1017.