Shapley-Study
Shapley-Study copied to clipboard
Continued training with input scores
trafficstars
I'm currently using LightGBM as the model for iterations of TMCShapley. LightGBM supports 'continued training with input scores', do you think it would be possible to tweak the implementation to significantly improve the speed of convergence on the exact DataShapley value?
Tried the above but think the cost of loading the model was large enough to neglect the performance gain of not training from scratch.