Currently, GBMs convert bipartite feature matrices to the SGSO format before predicting, requiring minimal code changes over the sklearn implementation. However, inference could go from $n_\text{pred}^2 \log (n_\text{train})$ to $n_\text{pred} \log (n_\text{train})$ with more clever code design (see the bipartite apply function in tree/_tree.pyx).
Currently, GBMs convert bipartite feature matrices to the SGSO format before predicting, requiring minimal code changes over the sklearn implementation. However, inference could go from $
n_\text{pred}^2 \log (n_\text{train})$ to $n_\text{pred} \log (n_\text{train})$ with more clever code design (see the bipartite apply function intree/_tree.pyx).