Fit Conditional Group DRO model in the target domain¶

In [ ]:

cgdro_(
    X_list, 
    y_list,
    X0 = NULL,
    index = NULL,
    family   = c("reg_ld", "reg_hd", "reg_ml", "cls"),
    f_learner = c("linear", "xgb", "xgb.cv", "high_d"),
    w_learner = c("logistic", "xgb", "xgb.cv", "ulsif"),
    loss_type = NULL,   # only used for reg_ld and reg_ml
    verbose = FALSE,
    ...
)

cgdro_( X_list, y_list, X0 = NULL, index = NULL, family = c("reg_ld", "reg_hd", "reg_ml", "cls"), f_learner = c("linear", "xgb", "xgb.cv", "high_d"), w_learner = c("logistic", "xgb", "xgb.cv", "ulsif"), loss_type = NULL, # only used for reg_ld and reg_ml verbose = FALSE, ... )

Arguments¶

X_list

List of feature matrices from each source (each n_l x p).

y_list

List of outcome vectors from each source (each n_l x 1).

X0

Target feature matrix (N x p); if NULL, use pooled source data.

index

index of the loading vector (1-based), the index-th coefficient is of interest (n_loading x 1). Only for family = 'reg_hd'.

family

family of CGDRO to solve, including 'reg_ld', 'reg_hd', 'reg_ml', 'cls'. More details.

f_learner

Outcome model learner. Options: "linear", "xgb", "xgb.cv", "high_d". Default is "xgb". If family = 'reg_ld' or family = 'reg_hd', we do not need to specify.

w_learner

Density ratio model learner. Options: "logistic", "xgb", "xgb.cv", "ulsif". Default is "logistic". If family = 'reg_ld' or family = 'reg_hd', we do not need to specify.

loss_type

Loss type for weight optimization. Options: "reward", "squaredloss", "regret". Default is "reward". Only needed for family = 'reg_ld' and family = 'reg_ml'.

verbose

Whether to print fitting progress (default: FALSE).

...

See family

Outputs¶

fit

A list of results including CGDRO aggregated weights (weight_), coefficient estimators (est_, est_bc_, est_plug_, coef_), and materials for later use in infer(), predict(), and summary().

Example¶

See family