Skip to content
Back to search results

Weighted PageRank

Introduced by
Xing, W. & Ghorbani, A.
Year
2004
Type
Centrality measure / ranking algorithm
Family
Network science
Font size

What does it do?

Weighted PageRank extends the classic PageRank algorithm to networks where edges carry different strengths or frequencies, distributing importance proportionally to both incoming and outgoing edge weights rather than treating all links equally.

This makes it substantially more informative than binary PageRank in any network where connection strength matters.

The intuition behind the method

Classic PageRank treats every hyperlink or tie identically: a node gains equal prestige from a faint connection as from a strong one. Weighted PageRank corrects this by scaling each vote by the relative weight of the edge. A node that receives many heavy-weight incoming ties from already-important nodes climbs highest in the ranking, while a node with numerous but weak connections remains modest. This mirrors how influence actually accumulates in most real-world networks, where interaction frequency or transaction volume signals true importance.

History

PageRank was introduced by Brin and Page in 1998 as the foundation of Google Search, treating all hyperlinks as equal votes. Weighted PageRank was formalised by Xing and Ghorbani in 2004, extending the algorithm to assign rank distribution proportional to edge weights, drawing on in-weight and out-weight fractions. Since then it has been applied across citation analysis, bioinformatics, social network analysis, and recommendation systems wherever link strength is a meaningful proxy for influence.

When to use it

Use Weighted PageRank when your network has meaningful edge weights and you need a global, flow-aware importance ranking of nodes. It is ideal for citation networks with co-citation counts, web graphs with traffic volumes, biological interaction networks with expression levels, social networks with interaction frequencies, and financial transaction networks. Prefer Weighted PageRank over binary PageRank whenever ignoring edge strength would misrepresent the flow of influence. Avoid it when edge weights are unreliable or arbitrarily assigned, when the network is very sparse and weights are mostly binary (plain PageRank suffices), or when you need a purely local centrality measure like weighted degree. It is not appropriate for undirected networks where flow direction is conceptually meaningless.

Limitations and considerations

  • Requires edge weights to be meaningful and reliably measured; noisy or arbitrary weights degrade ranking quality.
  • The damping factor (d) is a free parameter; the conventional 0.85 may not be optimal for non-web networks.
  • Sensitive to the weight normalisation scheme chosen; results can differ if in-weights and out-weights are combined differently.
  • Does not distinguish between different types of edges; a single-weight model may collapse important structural distinctions.
  • Convergence can be slow on very large, dense networks without efficient sparse-matrix implementations.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Incorporates edge strength directly into importance propagation, giving more realistic rankings than binary PageRank in weighted networks.
  • Captures global, network-wide prestige rather than only local connectivity.
  • Preserves the teleportation mechanism of PageRank, making it robust against dangling nodes and disconnected components when corrected.
  • Scales to large networks through efficient iterative computation; sparse matrix operations keep memory tractable.
  • Readily interpretable: a high score means a node receives heavy, weighted endorsement from other high-scoring nodes.

Applications

  • Identifying influential papers or authors in weighted citation and co-authorship networks.
  • Ranking web pages or documents when hyperlink traffic volume or click-through rates are available.
  • Detecting key proteins or genes in weighted biological interaction networks.
  • Prioritising nodes in financial transaction networks by transaction volume flow.
  • Ranking users or content in weighted social media interaction graphs.
  • Building recommendation systems where item co-occurrence frequencies serve as edge weights.

How it works

How it works

1

Represent the network as a graph where every edge carries a numerical weight reflecting tie strength — such as frequency of interaction, transaction volume, co-authorship count, or similarity score. The algorithm requires these weights to be non-negative; negative weights must be transformed before analysis.

2

For each directed edge from node v to node u, compute two weight factors: an in-weight factor normalising the edge weight by the total incoming weight of u, and an out-weight factor normalising it by the total outgoing weight of v. These fractions determine how much of v's rank is distributed to u via this specific edge.

Formula
W_{in}(v,u) = \frac{w(v,u)}{\sum_{k \in In(u)} w(k,u)}, \quad W_{out}(v,u) = \frac{w(v,u)}{\sum_{k \in Out(v)} w(v,k)}
3

Apply the Weighted PageRank formula iteratively: each node's rank equals a base teleportation term (1 minus damping factor d, typically 0.85) plus d times the sum over all in-neighbours of their rank scaled by the two weight factors. This distributes rank proportionally to edge strength rather than uniformly.

Formula
WPR(u) = (1-d) + d \sum_{v \in In(u)} WPR(v) \cdot W_{in}(v,u) \cdot W_{out}(v,u)
4

Repeat the rank update until the change in every node's score falls below a small tolerance (commonly 1e-6). Convergence is guaranteed for strongly connected graphs; for graphs with dangling nodes or isolated components, apply the standard dangling-node correction before iterating.

5

Once scores have converged, sort all nodes by their Weighted PageRank score in descending order. These ranks reflect each node's global structural importance accounting for both network topology and edge strength.

Common pitfalls

  • Using raw, un-normalised edge weights without checking their scale — mixing counts in the tens with values in the millions distorts rank propagation.
  • Forgetting to handle dangling nodes (nodes with no outgoing edges), which can drain rank from the entire network if uncorrected.
  • Assuming the default damping factor of 0.85 is universally appropriate; in citation or biological networks a different value may better reflect the domain.
  • Interpreting Weighted PageRank scores as absolute probabilities rather than relative rankings — only their order is stable across different runs.
  • Comparing Weighted PageRank scores across networks of different sizes without normalisation.

Frequently asked questions

How is Weighted PageRank different from standard PageRank?

Standard PageRank distributes a node's rank equally to all its out-neighbours regardless of edge strength. Weighted PageRank scales this distribution by the relative in-weight and out-weight of each edge, so stronger connections carry proportionally more rank.

What damping factor should I use?

The conventional value is 0.85, derived from web-graph assumptions about random surfing. For social, biological, or citation networks, values between 0.5 and 0.9 are explored in the literature. Sensitivity analysis over a range of damping values is recommended.

Can Weighted PageRank be applied to undirected networks?

Yes, by representing each undirected edge as a pair of directed edges with the same weight. The algorithm then produces symmetric in-weight and out-weight factors, which simplifies to a formula that still respects edge strength.

How do I handle zero-weight edges?

Edges with zero weight contribute nothing to rank propagation, so they can be removed before analysis. If zero weights are structurally meaningful (absence versus missing data), document the distinction clearly before removing them.

Is Weighted PageRank better than weighted degree centrality?

It depends on the research question. Weighted degree centrality captures only direct connections, while Weighted PageRank captures global prestige propagated across the whole network. Use weighted degree for local influence and Weighted PageRank for network-wide importance.

Related methods

Sources

  1. Xing, W., & Ghorbani, A. (2004). Weighted PageRank algorithm. Proceedings of the Second Annual Conference on Communication Networks and Services Research (CNSR '04), pp. 305–314. IEEE.

    Open DOI reference
    Copied.
  2. PageRank. Wikipedia.

    Open publication reference
    Copied.