PageRank and preferential attachment: why popularity can compound

Begin with nodes, links and the question being asked

A network model starts by defining what a node represents and what an edge means. A directed citation link differs from an undirected acquaintance. The same drawing can be used for different stories, but the interpretation of the calculations depends on that choice. Before explaining a hub, ask whether you are analysing an existing network or proposing a rule for how it formed.

The PageRank experiment fixes four nodes and lets you edit their directed links. The preferential-attachment experiment instead adds nodes and links over time. Its edges are undirected. Comparing them is useful precisely because they do not answer the same question: where does rank flow now, and which nodes receive new connections as the system grows?

PageRank follows probability mass rather than counting links

Brin and Page's original search-engine paper describes a prototype using web link structure. Modic's teaching model represents a surfer who follows an outgoing link with probability d and otherwise jumps uniformly. Rank received from a node depends on that node's own rank and how it distributes mass across its outgoing links.

A node with no outgoing links cannot simply swallow the probability. In this implementation its link-following mass is distributed uniformly. After one hundred iterations, the rank shares still sum to one. An incoming-link count would miss the source nodes' importance, their other outgoing links and the teleportation term. Those differences are visible even in a four-node example.

Worked example: edit one link, then remove link following

Use the small graph to test a mechanism rather than to derive a search-engine optimisation prescription. The exact rank percentages are conditional on this four-node network and its uniform jumps. Real web ranking includes far more structure and many signals absent from this model.

1. Inspect the initial graph: A links to B and C, B to C, C to A and D to C. Record D's rank and the total rank mass.
2. Add A→D using the link button. Compare D's new share and the shares that changed elsewhere; rank has been redistributed rather than created.
3. Set link-following probability to zero. Every node receives twenty-five percent because the process now consists entirely of uniform jumps, making the link graph irrelevant to this calculation.

Preferential attachment changes the future graph

Barabási and Albert's growing-network research combines growth with preferential attachment. In Modic's version, each new node adds one edge and selects an existing neighbour in proportion to current degree. A uniform alternative ignores degree. The comparison isolates a cumulative-advantage mechanism without assigning different intrinsic qualities to nodes.

An early connection advantage can create more opportunities to receive later connections. That is a mechanism for unequal degree, not proof that a hub is the best node. Nor does every unequal network arise this way. Capacity, deletion, copying, fitness and institutional rules can all change the process. The board contains forty nodes at most, so it cannot establish a power-law fit from its drawing.

A pattern does not uniquely identify its generating process

The power-law experiment explores concentration in a distribution. Preferential attachment offers one possible story for a related pattern of unequal connections, but a heavy tail alone does not identify that story. A useful explanation asks which observations distinguish the proposed mechanism from alternatives, rather than naming a famous distribution after seeing a few large values.

Likewise, a high rank on a fixed graph does not show why the graph formed. The system may already encode prior visibility, resource differences or selection. If ranking then changes what people encounter, the ranking can also influence future links. That feedback is a further model to investigate, not something automatically included in either of the two simple boards.

A practical audit of a popularity-based claim

When a ranking is used to allocate attention or resources, record what it measures and what people might mistakenly infer from it. A useful audit distinguishes structural prominence from the outcome the decision actually cares about. That distinction can reveal both the value and the limits of a ranking.

• Define nodes and the meaning and direction of each link.
• Separate an existing-graph calculation from a growth model.
• Check how dangling nodes and new nodes are treated.
• Ask whether source importance matters or only counts are used.
• Consider feedback from visibility to later opportunities.
• Do not substitute prominence for an independent quality measurement.

Keep search advice separate from a teaching model

The PageRank board does not reproduce a current search engine or provide a formula for traffic. Its small fixed graph omits content, queries, spam handling and many other features. Editing a link in it teaches mass flow; it does not establish what will happen if a real site gains one backlink. A credible practical claim would require relevant evidence about the actual system.

The durable lesson is about interpretation. Ask what signal the metric uses, how that signal was generated and what happens when people respond to it. Popularity can contain information and can also reflect cumulative opportunity. Understanding the mechanism helps you decide when prominence is useful evidence and when an independent check is needed before treating it as quality.