Convex Covariate Clustering for Classification

Python 3 Implementation of the method as proposed in http://arxiv.org/abs/1903.01680

Experiments on synthetic data

1. Training: Train with different hyperparameters, e.g.:

$ python syntheticDataExperiments.py smallFullContra 100 proposed

2. Run clustering evaluation: run all evaluation measures and save results, e.g.:

$ python evalAndSaveResults.py SYNTHETIC_DATA onlyNu 10 all smallFullContra 1000

3. Visualize Results: The results are save in the folder "plots".

$ python analyzeClustering.py SYNTHETIC_DATA smallFullContra 1000

ADMM Implementation

ADMM.py parallelADMM(dataFeatures, dataLabels, covariateSims, origNu, origGamma, optParams, warmStartAuxilliaryVars, paramId)

bStep.py optimizeLBFGSNew(dataFeatures, dataLabels, edgesZ, singleZ, edgesU, singleU, rho, B0, beta0, allAdjacentNodes, MAX_ITERATIONS)

zStep.py updateZ_fast(edgesZ, singleZ, edgesU, singleU, B, rho, S, nu, gamma, allAdjacentNodes)

uStep.py updateU_fast(edgesZ, singleZ, edgesU, singleU, B, allAdjacentNodes)

identify connected components: getClusteringAndSelectedFeatures(edgesU, singleU, B, rho, S, nu, gamma, allAdjacentNodesPrior) in zStep.py

Visualization Details

the visualization in "analyzeClustering.py" uses the colors extracted from: http://tools.medialab.sciences-po.fr/iwanthue/ and the graphviz package https://www.graphviz.org/