B3

---

This function determines the extrinsic clustering quality b-cubed measure using a set of known labels for a data set, and cluster assignmnets of each data point stored as either vector or cell arrays where each cell/row represets a data point in which an array containing class or cluster assignments is stored. The multi-class variant (.m) should not yet be used and does not scale well. The single-class variant is on the order of 10-1000 times faster than the other bcubed scoring available on pip, depending on the dataset, but the multiclass variant has not been implemented in python as it suffers from the same slow performace as bcubed.

Python and (matlab coming) implementations are provided.

Installation

Clone or download the repo. (I have to update the pip package)

Inputs

---

   L: An NxM matrix containing class labels for each data point (for the
      multiclass case, once implemented). Each row represents the ith data
      point and each column contains a 0, or 1 in the jth column indicating
      membership of label j. Alternatively, if hard class labels are
      available, L can be input as an Nx1 vector where each entry is its
      class label.

   K: Defined identically to L except for this variable stores cluster
      assignments for each data point

Outputs

---

  Fmeasure: This F-measure using the b-cubed metric
  precision: The b-cubed precision
  recall: The b-cubed recall

---

Performace demonstration

---

    ./compare\_performance.sh
    Timing shared setup time for 5 random cases using the bcubed package
    real    0m1.351s
    user    0m0.248s
    sys    0m0.071s

    Timing full execution time (b3 + shared) for the B3score implementation
    real    0m0.283s
    user    0m0.240s
    sys    0m0.045s

    Timing full execution time (bcubed + shared) for the bcubed implementation
    real    0m38.233s
    user    0m36.174s
    sys    0m0.273s

---

  Author: Matthew Wiesner
  Email : wiesner@jhu.edu 
  Institute: Johns Hopkins University Electrical and Computer Engineering

  Refences: 
       DESCRIPTION OF THE UPENN CAMP SYSTEM AS USED FOR COREFERENCE,
       Breck Baldwin, Tom Morton, Amit Bagga, Jason Baldridge, 
       Raman Chandraseker, Alexis Dimitriadis, Kieran Snyder, 
       Magdalena Wolska, Institute for Research in Cognitive Science

       A.A. Aroch-Villarruel Pattern Recognition 6th Mexican Conference,
       MCPR 2014 Proceedings Paper, p.115