Callback closure for collecting the model coefficients history of a gblinear booster during its training.
Callback closure for collecting the model coefficients history of a gblinear booster during its training.
cb.gblinear.history(sparse = FALSE)
sparse |
when set to FALSE/TURE, a dense/sparse matrix is used to store the result. Sparse format is useful when one expects only a subset of coefficients to be non-zero, when using the "thrifty" feature selector with fairly small number of top features selected per iteration. |
To keep things fast and simple, gblinear booster does not internally store the history of linear model coefficients at each boosting iteration. This callback provides a workaround for storing the coefficients' path, by extracting them after each training iteration.
Callback function expects the following values to be set in its calling frame:
bst (or bst_folds).
Results are stored in the coefs element of the closure.
The xgb.gblinear.history convenience function provides an easy way to access it.
With xgb.train, it is either a dense of a sparse matrix.
While with xgb.cv, it is a list (an element per each fold) of such matrices.
#### Binary classification:
#
# In the iris dataset, it is hard to linearly separate Versicolor class from the rest
# without considering the 2nd order interactions:
require(magrittr)
x <- model.matrix(Species ~ .^2, iris)[,-1]
colnames(x)
dtrain <- xgb.DMatrix(scale(x), label = 1*(iris$Species == "versicolor"))
param <- list(booster = "gblinear", objective = "reg:logistic", eval_metric = "auc",
lambda = 0.0003, alpha = 0.0003, nthread = 2)
# For 'shotgun', which is a default linear updater, using high eta values may result in
# unstable behaviour in some datasets. With this simple dataset, however, the high learning
# rate does not break the convergence, but allows us to illustrate the typical pattern of
# "stochastic explosion" behaviour of this lock-free algorithm at early boosting iterations.
bst <- xgb.train(param, dtrain, list(tr=dtrain), nrounds = 200, eta = 1.,
callbacks = list(cb.gblinear.history()))
# Extract the coefficients' path and plot them vs boosting iteration number:
coef_path <- xgb.gblinear.history(bst)
matplot(coef_path, type = 'l')
# With the deterministic coordinate descent updater, it is safer to use higher learning rates.
# Will try the classical componentwise boosting which selects a single best feature per round:
bst <- xgb.train(param, dtrain, list(tr=dtrain), nrounds = 200, eta = 0.8,
updater = 'coord_descent', feature_selector = 'thrifty', top_k = 1,
callbacks = list(cb.gblinear.history()))
xgb.gblinear.history(bst) %>% matplot(type = 'l')
# Componentwise boosting is known to have similar effect to Lasso regularization.
# Try experimenting with various values of top_k, eta, nrounds,
# as well as different feature_selectors.
# For xgb.cv:
bst <- xgb.cv(param, dtrain, nfold = 5, nrounds = 100, eta = 0.8,
callbacks = list(cb.gblinear.history()))
# coefficients in the CV fold #3
xgb.gblinear.history(bst)[[3]] %>% matplot(type = 'l')
#### Multiclass classification:
#
dtrain <- xgb.DMatrix(scale(x), label = as.numeric(iris$Species) - 1)
param <- list(booster = "gblinear", objective = "multi:softprob", num_class = 3,
lambda = 0.0003, alpha = 0.0003, nthread = 2)
# For the default linear updater 'shotgun' it sometimes is helpful
# to use smaller eta to reduce instability
bst <- xgb.train(param, dtrain, list(tr=dtrain), nrounds = 70, eta = 0.5,
callbacks = list(cb.gblinear.history()))
# Will plot the coefficient paths separately for each class:
xgb.gblinear.history(bst, class_index = 0) %>% matplot(type = 'l')
xgb.gblinear.history(bst, class_index = 1) %>% matplot(type = 'l')
xgb.gblinear.history(bst, class_index = 2) %>% matplot(type = 'l')
# CV:
bst <- xgb.cv(param, dtrain, nfold = 5, nrounds = 70, eta = 0.5,
callbacks = list(cb.gblinear.history(FALSE)))
# 1st forld of 1st class
xgb.gblinear.history(bst, class_index = 0)[[1]] %>% matplot(type = 'l')Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.