robyn_mmm: Core MMM Function

Description

robyn_mmm() function activates Nevergrad to generate samples of hyperparameters, conducts media transformation within each loop, fits the Ridge regression, calibrates the model optionally, decomposes responses and collects the result. It's an inner function within robyn_run().

Usage

robyn_mmm(
  InputCollect,
  hyper_collect,
  iterations,
  cores,
  nevergrad_algo,
  intercept = TRUE,
  intercept_sign,
  ts_validation = TRUE,
  add_penalty_factor = FALSE,
  objective_weights = NULL,
  dt_hyper_fixed = NULL,
  rssd_zero_penalty = TRUE,
  refresh = FALSE,
  trial = 1L,
  seed = 123L,
  quiet = FALSE,
  ...
)

Value

List. MMM results with hyperparameters values.

Arguments

InputCollect: List. Contains all input parameters for the model. Required when robyn_object is not provided.
hyper_collect: List. Containing hyperparameter bounds. Defaults to InputCollect$hyperparameters.
iterations: Integer. Number of iterations to run.
cores: Integer. Default to parallel::detectCores() - 1 (all cores except one). Set to 1 if you want to turn parallel computing off.
nevergrad_algo: Character. Default to "TwoPointsDE". Options are c("DE","TwoPointsDE", "OnePlusOne", "DoubleFastGADiscreteOnePlusOne", "DiscreteOnePlusOne", "PortfolioDiscreteOnePlusOne", "NaiveTBPSA", "cGA", "RandomSearch").
intercept: Boolean. Should intercept(s) be fitted (default=TRUE) or set to zero (FALSE).
intercept_sign: Character. Choose one of "non_negative" (default) or "unconstrained". By default, if intercept is negative, Robyn will drop intercept and refit the model. Consider changing intercept_sign to "unconstrained" when there are context_vars with large positive values.
ts_validation: Boolean. When set to TRUE, Robyn will split data by test, train, and validation partitions to validate the time series. By default the "train_size" range is set to c(0.5, 0.8), but it can be customized or set to a fixed value using the hyperparameters input. For example, if train_size = 0.7, validation size and test size will both be 0.15 and 0.15. When ts_validation = FALSE, nrmse_train is the objective function; when ts_validation = TRUE, nrmse_val is the objective function.
add_penalty_factor: Boolean. Add penalty factor hyperparameters to glmnet's penalty.factor to be optimized by nevergrad. Use with caution, because this feature might add too much hyperparameter space and probably requires more iterations to converge.
objective_weights: Numeric vector. Default to NULL to give equal weights to all objective functions. Order: NRMSE, DECOMP.RSSD, MAPE (when calibration data is provided). When you are not calibrating, only the first 2 values for objective_weights must be defined, i.e. set c(2, 1) to give double weight to the 1st (NRMSE). This is an experimental feature. There's no research on optimal weight setting. Subjective weights might strongly bias modeling results.
dt_hyper_fixed: data.frame or named list. Only provide when loading old model results. It consumes hyperparameters from saved csv pareto_hyperparameters.csv or JSON file to replicate a model.
rssd_zero_penalty: Boolean. When TRUE, the objective function DECOMP.RSSD will penalize models with more 0 media effects additionally. In other words, given the same DECOMP.RSSD score, a model with 50% 0-coef variables will get penalized by DECOMP.RSSD * 1.5 (larger error), while another model with no 0-coef variables gets un-penalized with DECOMP.RSSD * 1.
refresh: Boolean. Set to TRUE when used in robyn_refresh().
trial: Integer. Which trial are we running? Used to ID each model.
seed: Integer. For reproducible results when running nevergrad.
quiet: Boolean. Keep messages off?
...: Additional parameters passed to robyn_outputs().