Collaborative knowledge systems evolve through messy, user-generated hierarchies. This study proposes a centrality-based pruning method that cleanses the Wikipedia category network by identifying structural redundancies and inconsistencies. By relying solely on endogenous information, it demonstrates how collective curation and algorithmic structure can be combined to improve the organization of open knowledge — work that later became the basis for a Wikipedia-based benchmark for graph neural networks, now widely used in research.

Discovering non-obvious relations in knowledge systems requires models that go beyond surface similarity. This paper presents an online margin-based learning algorithm that infers a latent matrix of category interactions to uncover hidden connections within Wikipedia’s hyperlink structure. The method efficiently scales to large graphs, revealing how semantic structures and user-generated organization jointly shape information discovery.