Did DMatrix initialized with dense or sparse matrix lead to different result

[breakhearts]

I tried both R and python package. When i initialized DMatrix with dense

dmodel = xgb.DMatrix(model, label=Y.iloc[40000:].values, feature_names=dt.columns)

or sparse

dmodel = xgb.DMatrix(csc_matrix(model), label=Y.iloc[40000:].values, feature_names=dt.columns)

The training accuracy is different, is that correct? And should I always use sparse matrix?

[khotilov]

The training accuracy is different, is that correct?

Yes. "Sparse" elements are treated as "missing" by the tree booster and as zeros by the linear booster.

And should I always use sparse matrix?

Use whichever works better for you.

[adamwlev]

This was totally unexpected for me. Thanks for raising this @breakhearts