Hero Research Group
menu MENU

AFFECT MATLAB Toolbox for clustering dynamic data

Kevin S. Xu, Mark Kliger, Alfred O. Hero III

Version 0.6 (released February 28, 2014)

Download (.zip)

Currently included in the toolbox:

  • Batch implementations of AFFECT with k-means, hierarchical, and spectral clustering
  • Demos of AFFECT with k-means, hierarchical, and spectral clustering on two colliding Gaussians simulated data
  • Demo of AFFECT with spectral clustering on MIT Reality Mining data

The AFFECT (Adaptive Forgetting Factor for Evolutionary Clustering and Tracking) MATLAB toolbox is designed for clustering dynamic data sets and tracking communities in dynamic networks.

Please refer to the paper, available at http://arxiv.org/abs/1104.1990, for additional details.

Changelog

Version 0.6

  • Updated calculation of sample mean and variance used in estimation of optimal forgetting factor to be much faster (noticeable when clustering 1,000 objects)

Version 0.5

  • Added batch implementations of AFFECT with k-means and hierarchical clustering
  • Changed data structure for proximity matrices; object IDs (names) are now kept in a separate cell array from the proximity matrices
  • Corrected an object ordering bug that produced incorrectly ordered clustering results when object IDs were not sorted
  • Corrected a bug in permute_clusters_opt.m that matched clusters suboptimally when the number of clusters changed over time
  • Added demos of AFFECT k-means, hierarchical, and spectral clustering on two colliding Gaussians simulated data
  • Updated demo of AFFECT on MIT Reality Mining data
  • Significantly expanded Readme file

Version 0.2:

  • Fixed a bug that prevented the modularity and silhouette cluster selection heuristics from being used
  • Created two options for eigengap heuristic: one that prompts the user to pick the number of clusters only once (the number is kept constant over all time steps) and one that prompts to pick the number of clusters at each iteration in each time step (annoying but allows the number of clusters to vary over time)
  • Added functions modularity.m and parse_inputs.m that were mistakenly not included in version 0.1
  • Added recommended MATLAB toolboxes and descriptions of files to the Readme

Version 0.1:

  • Initial release containing batch implementation of AFFECT with spectral clustering and demo on MIT Reality Mining data

Contact