Grow Langevin Trees — grow_langevin

CatBoost Stochastic Gradient Langevin Boosting (SGLB): injects calibrated Langevin diffusion noise into gradient updates, converting boosting into a Markov chain sampler over function space. Guarantees global convergence for non-convex losses where standard GBM can only guarantee local optima. See https://arxiv.org/abs/2001.07248.

Usage

grow_langevin_trees(
  x,
  y,
  trees = 100L,
  task = "classification",
  diffusion_temperature = 10000,
  model_shrink_rate = 0.05,
  exotic = FALSE,
  ...
)

Arguments

x: a data X
y: a data Y
trees: num of boosting rounds
task: a task
diffusion_temperature: Langevin noise amplitude; higher = more exploration
model_shrink_rate: shrinkage applied to older trees each round
exotic: logical; if TRUE, apply additional exotic parameters
...: other eng args

Value

a fitted model

Details

When exotic = FALSE (default), parameters follow the SGLB paper: Langevin mode with diffusion temperature and model shrink rate, Lossguide growth, and iterated Newton leaf estimation.

When exotic = TRUE, second-order split scoring (NewtonL2), Bayesian bootstrap row weighting, and increased split candidate density (border_count) are added — treating the Langevin chain as a proper Bayesian posterior sampler over tree space.