Each round fits, for every class, a least-squares tree to y_ik - p_k(x_i)
and sets each leaf to leaf_scale sum r / sum |r| (1 - |r|) (ESL Alg. 10.4, eq
10.57, leaf_scale = (K - 1) / K; gbm's multinomial uses 1), shrunk by nu;
p_k is the softmax of the K functions (eq 10.21).
Usage
morie_esl_gbm_multiclass(
X,
g,
M = 100,
nu = 0.1,
max_depth = 1,
min_leaf = 1,
query = NULL,
leaf_scale = NULL
)
Arguments
- X
Predictors.
- g
Class labels.
- M
Rounds.
- nu
Shrinkage.
- max_depth, min_leaf
Tree controls.
- query
Points to classify (default X).
- leaf_scale
Newton step factor (default (K - 1) / K).
Value
Named list: prob, prediction, deviance_path, classes, trees.
References
Friedman, J. (2001). Annals of Statistics 29, 1189-1232.
Examples
i <- 1:60
X <- cbind(sin(i) + ((i %% 3) == 0) * 1.5, cos(2 * i))
morie_esl_gbm_multiclass(X, i %% 3, M = 5)$deviance_path
#> [1] 126.0375 121.4441 117.7253 114.3567 111.4577
