| audrex {audrex} | R Documentation |
audrex: Automatic Dynamic Regression using Extreme Gradient Boosting
Description
Dynamic regression for time series using Extreme Gradient Boosting with hyper-parameter tuning via Bayesian Optimization or Random Search.
Usage
audrex(
data,
n_sample = 10,
n_search = 5,
smoother = FALSE,
seq_len = NULL,
diff_threshold = 0.001,
booster = "gbtree",
norm = NULL,
n_dim = NULL,
ci = 0.8,
min_set = 30,
max_depth = NULL,
eta = NULL,
gamma = NULL,
min_child_weight = NULL,
subsample = NULL,
colsample_bytree = NULL,
lambda = NULL,
alpha = NULL,
n_windows = 3,
patience = 0.1,
nrounds = 100,
dates = NULL,
acq = "ucb",
kappa = 2.576,
eps = 0,
kernel = list(type = "exponential", power = 2),
seed = 42
)
Arguments
data |
A data frame with time features on columns. |
n_sample |
Positive integer. Number of samples for the Bayesian Optimization. Default: 10. |
n_search |
Positive integer. Number of search steps for the Bayesian Optimization. When the parameter is set to 0, optimization is shifted to Random Search. Default: 5, |
smoother |
Logical. Perform optimal smoothing using standard loess. Default: FALSE |
seq_len |
Positive integer. Number of time-steps to be predicted. Default: NULL (automatic selection) |
diff_threshold |
Positive numeric. Minimum F-test threshold for differentiating each time feature (keep it low). Default: 0.001. |
booster |
String. Optimization methods available are: "gbtree", "gblinear". Default: "gbtree". |
norm |
Logical. Boolean flag to apply Yeo-Johson normalization. Default: NULL (automatic selection from random search or bayesian search). |
n_dim |
Positive integer. Projection of time features in a lower dimensional space with n_dim features. The default value (NULL) sets automatically the values in c(1, n features). |
ci |
Confidence interval. Default: 0.8. |
min_set |
Positive integer. Minimun number for validation set in case of automatic resize of past dimension. Default: 30. |
max_depth |
Positive integer. Look to xgboost documentation for description. A vector with one or two positive integer for the search boundaries. The default value (NULL) sets automatically the values in c(1, 8). |
eta |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric between (0, 1] for the search boundaries. The default value (NULL) sets automatically the values in c(0, 1). |
gamma |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric for the search boundaries. The default value (NULL) sets automatically the values in c(0, 100). |
min_child_weight |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric for the search boundaries. The default value (NULL) sets automatically the values in c(0, 100). |
subsample |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric between (0, 1] for the search boundaries. The default value (NULL) sets automatically the values in c(0, 1). |
colsample_bytree |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric between (0, 1] for the search boundaries. The default value (NULL) sets automatically the values in c(0, 1). |
lambda |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric for the search boundaries. The default value (NULL) sets automatically the values in c(0, 100). |
alpha |
Positive numeric. Look to xgboost documentation for description. A vector with one or two positive numeric for the search boundaries. The default value (NULL) sets automatically the values in c(0, 100). |
n_windows |
Positive integer. Number of (expanding) windows for cross-validation. Default: 3. |
patience |
Positive numeric. Percentage of waiting rounds without improvement before xgboost stops. Default: 0.1 |
nrounds |
Positive numeric. Number of round for the extreme boosting machine. Look to xgboost for description. Default: 100. |
dates |
Date. Vector of dates for the time series. Default: NULL (progressive numbers). |
acq |
String. Parameter for Bayesian Optimization. For reference see rBayesianOptimization documentation. Default: "ucb". |
kappa |
Positive numeric. Parameter for Bayesian Optimization. For reference see rBayesianOptimization documentation. Default: 2.576. |
eps |
Positive numeric. Parameter for Bayesian Optimization. For reference see rBayesianOptimization documentation. Default: 0. |
kernel |
List. Parameter for Bayesian Optimization. For reference see rBayesianOptimization documentation. Default: list(type = "exponential", power = 2). |
seed |
Random seed. Default: 42. |
Value
This function returns a list including:
history: a table with the models from bayesian (n_sample + n_search) or random search (n_sample), their hyper-parameters and optimization metric, the weighted average rank
models: a list with the details for each model in history
best_model: results for the best selected model according to the weighted average rank, including:
predictions: min, max, q25, q50, q75, quantile at selected ci, mean, sd, skewness and kurtosis for each time feature
joint_error: max sequence error for the differentiated time features (max_rmse, max_mae, max_mdae, max_mape, max_mase, max_rae, max_rse, max_rrse, both for training and testing)
serie_errors: sequence error for the differentiated time features averaged across testing windows (rmse, mae, mdae, mape, mase, rae, rse, rrse, both for training and testing)
pred_stats: for each predicted time feature, IQR to range, divergence, risk ratio, upside probability, averaged across prediction time-points and at the terminal points
plots: a plot for each predicted time feature with highlighted median and confidence intervals
time_log
Author(s)
Giancarlo Vercellino giancarlo.vercellino@gmail.com
See Also
Useful links:
Examples
audrex(covid_in_europe[, 2:5], n_samp = 3, n_search = 2, seq_len = 10) ### BAYESIAN OPTIMIZATION
audrex(covid_in_europe[, 2:5], n_samp = 5, n_search = 0, seq_len = 10) ### RANDOM SEARCH