Skip to content

Latest commit

 

History

History
8 lines (7 loc) · 608 Bytes

File metadata and controls

8 lines (7 loc) · 608 Bytes

Forest Garrote training:

  • The sklearn Lars algorithm does not accept sparse input, so it cannot be used.
  • If you compute the SVM weights instead of the Lasso weights, the resulting node weights are far from sparse.
  • Precomputing the gram matrix for large trees is (a) computationally expensive and (b) memory consuming, since the result is of shape (num_nodes, num_nodes) and probably not sparse. Therefore the trees are gathered in small groups (4-5 trees) and the forest garrote is computed on each group.

Literature / People:

  • Lawson / Hanson: Solving Least Squares Problems
  • Philipp Gähring