HITS Algorithm¶
A link-analysis algorithm that iteratively assigns each node an authority score from incoming high-hub links and a hub score from outgoing links to high-authority nodes, usually on a query-dependent directed subgraph.
Core Idea¶
HITS treats endorsement and curation as different graph roles. Authorities receive links from hubs, and hubs gain value by pointing to authorities, so the two score vectors define one another.
The elegance of the eigensystem does not eliminate data choices. Query set, graph expansion, link direction, duplicates, spam, weighting, normalization, and convergence determine what the final ranking means.
Scope of Application¶
- Web search research. Ranks topical pages by link roles.
- Citation networks. Separates review-like hubs from cited authorities cautiously.
- Knowledge graphs. Finds mutually reinforcing source and target roles.
- Spam analysis. Examines how coordinated linking distorts spectral scores.
Clarity¶
Report node and edge definition, adjacency orientation, query/root/base construction, edge weighting, preprocessing, initialization, normalization, tolerance, iteration count, handling of disconnected components, score scaling, and relevance or manipulation evaluation. Inclusion test: Require separate mutually reinforcing hub and authority scores derived iteratively from a directed link matrix or an explicitly equivalent eigensystem. Exclusion test: Exclude raw in/out-degree ranking, PageRank's single random-walk importance score, keyword relevance alone, and any two-score classifier without link reinforcement. Nearest boundary: PageRank distributes one prestige score through outgoing links with damping; HITS separates source-like hubs from target-like authorities and is often run on a topical subgraph. Exit condition: The output stops supporting the intended ranking when graph extraction, edge direction, weighting, normalization, convergence, or query expansion is unspecified or adversarially distorted. Common misclassifications: HITS is not simple link counting. Hub and authority scores are not interchangeable. It is not PageRank. High score does not independently verify truth or quality. Nearest named distinctions: PageRank: Produces one damped random-walk prestige score. Eigenvector centrality: Usually assigns one recursively reinforced node score. Indegree: Counts incoming edges without source-quality weighting. SALSA: Combines hub/authority ideas with stochastic walks.
Manages Complexity¶
Two short matrix recurrences compress a directed graph into complementary rankings. Spectral simplicity makes results sensitive to graph boundaries, dense communities, and the social meaning assigned to links.
Abstract Reasoning¶
- Construct the directed graph and document source-to-target orientation.
- If query-focused, build root and expanded base sets under explicit rules.
- Initialize hub and authority vectors and alternate matrix updates.
- Normalize each iteration and stop under a declared convergence criterion.
- Inspect dominant communities, topic drift, spam, and stability before interpreting scores.
Knowledge Transfer¶
Mutual source/target reinforcement transfers to citation and recommendation graphs, but the semantics of an edge and the legitimacy of query expansion must be rebuilt. A hyperlink vote is not automatically analogous to trust or truth.
Relationships to Other Abstractions¶
Current abstraction HITS Algorithm Domain-specific
Parents (1) — more general patterns this builds on
-
HITS Algorithm is a kind of Search and Retrieval Prime
HITS Algorithm is a strict kind of Search and Retrieval: it iteratively ranks hubs and authorities to retrieve structurally relevant web nodes.
Hierarchy paths (4) — routes to 3 parentless roots
- HITS Algorithm → Search and Retrieval → Problem Space → Representation → Abstraction
- HITS Algorithm → Search and Retrieval → Trade-offs → Constraint
- HITS Algorithm → Search and Retrieval → Problem Space → State and State Transition → Phase Space
- HITS Algorithm → Search and Retrieval → Problem Space → Problem Representation → Representation → Abstraction
Neighborhood in Abstraction Space¶
HITS Algorithm sits in a moderately populated region (42nd percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Computer Systems & Network Architecture (20 abstractions)
Nearest neighbors
- Blockmodel — 0.87
- Prim’s Algorithm — 0.87
- Routing — 0.87
- Commons-Based Peer Production — 0.87
- Tensor Network — 0.87
Computed from structural-signature embeddings · 2026-10-08