Python Implementation of Weighted Generalized Canonical Correlation Analysis as described in "Learning Multiview Embeddings of Twitter Users". Benton A, Arora R, and Dredze M. ACL 2016.
Tested with
- Python 2.7
- scipy 0.17.0
- numpy 1.10.4
Test suite:
python src/wgccaTest.py
Sample call to learn 5-dimensional WGCCA model (first two views weighted twice as much as second two):
python src/wgcca.py --input resources/sample_wgcca_input.tsv.gz --output wgcca_embeddings.npz --model wgcca_model.pickle --k 5 --kept_views 0 1 2 3 --weights 1.0 1.0 0.5 0.5 --reg 1.e-8 1.e-8 1.e-8 1.e-8
- Input format can be grokked from:
resources/sample_wgcca_input.tsv
- WGCCA model saved to:
wgcca_model.pickle
- WGCCA embeddings saved to:
wgcca_embeddings.npz
_compute
: look at this if you want to know how embeddings are computedlearn
: entrypoint for learning WeightedGCCA model from training setapply
: entrypoint for extracting embeddings from new data
The input views used in "Learning Multiview Embeddings of Twitter Users" can be found at http:https://www.cs.jhu.edu/~mdredze/datasets/multiview_embeddings/ -- in the same format as resources/sample_wgcca_input.tsv
.
If you use this code please cite:
Adrian Benton, Raman Arora, and Mark Dredze. Learning Multiview Representations of Twitter Users. Association for Computational Linguistics (ACL), 2016.
Please contact adrian dot author1_surname at gmail dot com if you have any questions/suggestions/concerns/comments.