HITS Algorithm

The core logic of HITS is a mutually reinforcing relationship between hubs and authorities. A good hub points to many good authorities, and a good authority is pointed to by many good hubs. This creates a recursive definition where the quality of one node type depends directly on the quality of the other. The algorithm iteratively updates hub and authority scores until convergence, effectively capturing the latent semantic structure of the web graph without requiring a query-independent global ranking.

Unlike PageRank, which computes a single global importance score for every page offline, HITS is a query-time algorithm. It first constructs a focused subgraph by collecting a root set of pages matching the query and expanding it with their immediate neighbors. Authority and hub scores are then calculated exclusively within this local neighborhood. This makes HITS highly specific to the query topic but computationally expensive to run at scale for every search request.

A significant limitation of HITS is topic drift, where the expanded subgraph inadvertently includes pages that are popular but off-topic. If a highly connected page about a different subject is included in the neighborhood, its strong link structure can dominate the iterative calculation, causing the top-ranked authorities to drift away from the original query intent. This phenomenon is a direct consequence of the local graph expansion step and requires careful filtering to mitigate.

The algorithm is susceptible to the Tightly-Knit Community (TKC) effect, where a small, densely interlinked cluster of pages on a subtopic can monopolize the top authority scores, even if they represent a minority viewpoint. Because the cluster's internal links create a strong mutual reinforcement loop, it can outscore more globally relevant but less densely connected authorities. This is a structural bias inherent to the eigenvector-based calculation within a local graph.

Mathematically, HITS is an eigenvector calculation on the adjacency matrix of the focused subgraph. The authority scores converge to the principal eigenvector of the matrix AᵀA, where A is the adjacency matrix, while the hub scores converge to the principal eigenvector of AAᵀ. This formulation reveals that authorities are the most referenced pages within the dominant theme of the local link structure, and hubs are the pages that reference them most comprehensively.