Sparse matrices w/ sklearn api

[dagley11]

I'm getting very different scores when using a scipy.sparse.csr_matrix as an input vs a Pandas Data Frame for XGBRegressor.fit() I checked this behavior against scikit-learn using random forests and my scores did not deviate, so the issue is not with the sparse matrix. Any ideas?

[tqchen]

xgboost is optimized for sparse matrix treating non-presenting entries as misisng, while sklearn does not. and it really depends on what preprocessing was done to pandas data frame.

[hlbkin]

Having same issue. I have sparse dataset of user clicks on websites, and when I fit XGBclassifier (xgb sklearn wrapper) with src matrix I get way worse results than if doing with standard numpy.

My data is basically 1/0 (visited/nonvisited).

I still do not understand why results differ so much

[bryan-woods]

This was also noted in #1238. There is definitely a problem with sparse matrixes in the latest sklearn API.

[tqchen]

#1583