Examples

Ensembles by changing K

• Use Majority Vote to stabilize k-means clustering
  • Example of calculating and plotting a single clustering solution
  • Build the ensemble of many heuristic k-means solutions with large K
  • Calculate and plot two validation metrics across majority voting solutions
  • Plot the co-occurrence matrix of the ensemble
• Emergent property of ensembles - uncovering different structure types than parent algorithm (Kmeans, Majority Vote, half moons)
  • Example of using OpenEnsembles to create and visualize a single solution
  • Create an ensemble of kmeans, plot convergence towards solution

Ensembles by changing the random seed

• Demonstrate the ability to get reproducible results from non-deterministic algorithms
  • Compare results with and without using a random seed: Kmeans
    • Mutual information should show results vary when seed is not forced to the same starting point
    • When the seed is forced to the same, Kmeans should return the same results, as indicated by mutual information of 1 between all clustering results.
  • Compare results with and without random seed for spectral clustering
    • Mutual information should show results vary when seed is not forced to the same starting point
    • When the seed is forced to the same, Spectral clustering should return the same results, as indicated by mutual information of 1 between all clustering results.

Ensembles by changing Distances and/or Algorithms

• Build an ensemble using algorithms and distances, finish to see if it uncovers the lack of structure
  • Create an ensemble
  • Plot example solutions from the ensemble
  • Finish the ensemble using Majority Vote and Graph Closure
  • Use mutual information to compare solutions
• Some algorithms take distance or similarity as a parameter. Here we demonstrate their effects on the solution space.
  • Demonstration of varying distance for algorithm based on distance metrics
  • Demonstration of varying distance for algorithm based on affinity metrics

Ensembles by subsampling data features

• Demonstration of subsampling feature dimensions
  • How sensitive is feature subsampling to transforming before or after sampling?
  • Finish the ensembles using the co-occurrence matrix

Ensembles that account for experimental noise

• Create resampling of available replicate data to explore effects of noise
  • Cluster all resampled data for an ensemble solution

Ensembles that use parallelization

• Parallelization of cluster and validation: Kmeans, Majority Vote on half moons
  • Example of using OpenEnsembles to create and visualize a single solution
  • Create an ensemble of kmeans, plot convergence towards solution with parallelization
• Parallelization of data and cluster objects: Resampling data to explore effects of noise
  • Cluster all resampled data for an ensemble solution