BCubed

BCubed is a family of evaluation metrics for clustering and classification that are designed to handle overlapping clusters. It was introduced by Pushpak Bagga and Andrew Baldwin in 1998 to evaluate entity-based cross-document clustering. BCubed defines precision and recall at the item level and then aggregates these scores across all items to produce a single overall measure, typically reported as an F1 score. A defining feature of BCubed is its ability to accommodate items that belong to multiple clusters or multiple true categories.

To compute BCubed precision, for each item i, consider all predicted clusters that contain i. For each such cluster, compute the proportion of items in that cluster that share at least one gold label with i, and then average these proportions across all clusters containing i. To compute BCubed recall, for each item i, consider all gold clusters that contain i and compute the proportion of items in those gold clusters that share at least one predicted cluster with i, then average across i. The overall BCubed precision and recall are the averages of these per-item values, and the BCubed F1 is the harmonic mean of precision and recall.

BCubed is widely used in information retrieval and natural language processing to evaluate document clustering, named-entity disambiguation, and coreference resolution tasks. It is praised for its suitability to overlapping clusters and its intuitive interpretation, but it has also been criticized for sensitivity to label distribution and for occasional counterintuitive results in certain edge cases. Variants and extensions exist to adapt BCubed to specific applications and data characteristics.