On Sat, Nov 12, 2011 at 12:19:37AM +0900, Mathieu Blondel wrote: > GraphLasso seems really neat (and the associated CV object should > prove very useful).
Thanks! > I see that the (dense) correlation matrix is used as input to affinity > propagation. Wouldn't it be better if we used the partial correlations > learned by GraphLasso directly? I don't like the results as much. Basically, we can either cluster on conditional relations, or on marginal relations: basically, as an input of affinity propagation, we can use the correlation matrix, which is a standard affinity measure, or the precision matrix, which is more like a condition affinity. They will give different information. Both information are interesting. That said, we can run the affinity propagation on the covariance matrix estimated by the GraphLasso, it gives the same result, but may be more clear in terms of linking the model. If people think it gives more clear picture, I'll implement the change. > This way, the cluster centers and the edge structure would be more > related. I actually like the fact that they give different information. > Likewise, the graph structure and the clustering are not used at all > for learning the embedding, which seems like a pity. I struggled a bit to use the graph structure in different embeddings techniques. For instance with the kernel PCA, or some spectral embedding. It didn't give good results, so I moved along. > Also, I wonder if once we have the graph structure, couldn't we get > away with just using a graph drawing algorithm? We don't have any :). But anyhow, a graph drawing is really an embedding learning algorithm, that takes an affinity matrix, so I don't see much difference. Gael ------------------------------------------------------------------------------ RSA(R) Conference 2012 Save $700 by Nov 18 Register now http://p.sf.net/sfu/rsa-sfdev2dev1 _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
