zygmuntz / sofia-ml-mod

sofia-kmeans with sparse RBF cluster mapping

sofia-ml, modified

This is a modified version of sofia-ml by D. Sculley. It differs from the original only in two additional parameters for sofia-kmeans RBF cluster mapping:

--cluster_mapping_threshold <float t>

After mapping, the values below the threshold will be set to zero (try t = 0.01 or 0.001). The resulting file might be smaller and sparsifying in this way may also help supervised learning (smaller error in classification or regression).

Make sure that t is small enough in relation to --cluster_mapping_param, because if all the features are smaller than the threshold, you'll just get a bunch of zeros.

--cluster_mapping_sparsity <int s>

Only use s nearest cluster centers for mapping. It is a stronger alternative to setting a threshold. The results with sparse mapping might be slightly worse, but the file certainly will be way smaller. Recommended for large data sets.

The modified files are:

sofia-ml/cluster-src/sf-cluster-centers.cc
sofia-ml/cluster-src/sf-cluster-centers.h
sofia-ml/src/sofia-kmeans.cc

See http://fastml.com/the-secret-of-the-big-guys/ for more.

sofia-ml-mod's Issues

SGD-SVM Parameters

Hello,
i'm new to all of this and i apologize for this question.

I'm using the Sofia-ML SGD-SVM, as i understand it is the same like RankSVM but with SGD. Can you maybe explain me how parameters like:

Learning Options:
-c float -> C: trade-off between training error
and margin (default 0.01)
Kernel Options:
-t int -> type of kernel function:

are set in Sofia-ML? Are they set to default?

Thanks and Regards