A sequential pipeline of model layers

mb_seq chains one or more mb_layer objects together. Each layer runs after the previous one, receiving a dataset enriched with all predictions produced so far. This allows later layers to use earlier predictions as input features.

Usage

mb_seq(
  ...,
  train = NULL,
  test = NULL,
  holdout = NULL,
  y_name = NA_character_,
  expo_name = NA_character_,
  model_display_name = NA_character_
)

Arguments

...

One or more mb_layer objects, in execution order.

train, test, holdout

Data frames for model development and evaluation.

y_name

[character(1)] Final target variable name. Used by diagnostic functions.

expo_name

[character(1)] Exposure column name. Default NA.

model_display_name

[character(1)] Human-readable label.

Value

An mb_seq object.

Details

At construction time, mb_seq validates that every blueprint's required columns (@x_original_inputs, @expo_name, @offset_name) are present in the supplied data at the point that blueprint would run – accounting for the fact that earlier layers add new columns.

See also

mb_layer() to construct individual layers, predict.mb_seq() for generating predictions.

Examples

if (FALSE) { # \dontrun{
# Pure premium: frequency * severity
seq_ps <- mb_seq(
  mb_layer(
    blueprints = list(mb_freq, mb_sev),
    yhat_name  = "pred_pure",
    aggregate_fn     = function(df) {
      df[["pred_pure"]] <- df[["pred_freq"]] * df[["pred_sev"]]
      df
    }
  ),
  train              = df_train,
  y_name             = "burn_cost",
  expo_name          = "earned_premium",
  model_display_name = "pure_premium"
)

# Expected loss: PD x EAD x LGD
seq_el <- mb_seq(
  mb_layer(
    blueprints = list(mb_pd, mb_ead, mb_lgd),
    yhat_name  = "pred_el",
    aggregate_fn     = function(df) {
      df[["pred_el"]] <- df[["pred_pd"]] * df[["pred_ead"]] * df[["pred_lgd"]]
      df
    }
  ),
  train  = df_train,
  y_name = "actual_loss"
)

# Sequential: GLM output feeds XGBoost as a feature
seq_chain <- mb_seq(
  mb_layer(list(mb_glm)),
  mb_layer(list(mb_xgb)),
  train  = df_train,
  y_name = "target"
)
} # }