WeightedBlender

class lightautoml.automl.blend.WeightedBlender(max_iters=5, max_inner_iters=7, max_nonzero_coef=0.05)[source]

Bases: lightautoml.automl.blend.Blender

Weighted Blender based on coord descent, optimize task metric directly.

Weight sum eq. 1. Good blender for tabular data, even if some predictions are NaN (ex. timeout). Model with low weights will be pruned.

__init__(max_iters=5, max_inner_iters=7, max_nonzero_coef=0.05)[source]
Parameters

  • max_iters (int) – Max number of coord desc loops.

  • max_inner_iters (int) – Max number of iters to solve inner scalar optimization task.

  • max_nonzero_coef (float) – Maximum model weight value to stay in ensemble.