LasCCA with Support for Missing Views
Python implementation of Large-scale GCCA (LasCCA) adapted to support data with no active features in views. LasCCA is described here:
Fu, Xiao, et al. "Efficient and distributed algorithms for large-scale generalized canonical correlations analysis." Data Mining (ICDM), 2016 IEEE 16th International Conference on. IEEE, 2016.
Run synthetic validation experiments with:
python synth_gcca.py
Sample input data in sample_data/N-20000_M-100_V-3_rho-1.000000e-01_sparsity-0.000000e+00.train.npz
(compressed numpy array). Train a model with:
python scalable_gcca.py --in_path ../sample_data/N-20000_M-100_V-3_rho-1.000000e-01_sparsity-0.000000e+00.train.npz --view 0 1 2 -k 5 --epochs 5 --max_cg_iters 20 --warmstart_cg --prop_heldout 0.0 --model_path ./gcca_retrained_N-20000_M-100_V-3_rho-1.000000e-01_sparsity-0.000000e+00.model.npz --projected_path N-20000_M-100_V-3_rho-1.000000e-01_sparsity-0.000000e+00.projected.npz
Thanks to Xiao Fu for a Matlab reference LasCCA implementation: http:https://people.oregonstate.edu/~fuxia/