mediamix: media transforms and attribution path construction

Feature engineering for marketing mix modelling and multi-touch attribution. mediamix turns raw marketing data into model-ready features: media spend into carryover- and saturation-adjusted regressors, and raw event logs into attributed customer journeys.

Details

It is a preprocessing and reporting package. It does not fit models – lm(), glmnet and brms do that better than a marketing package would. It computes order-1 Markov removal effects natively; for higher-order chains as_channel_paths() hands your journeys to ChannelAttribution.

Full documentation, an end-to-end walkthrough and the methods behind each function are at https://elkronos.github.io/mediamix/.

Where to start

Five vignettes, in the order most people need them:

vignette("mediamix"): Getting started: raw weekly spend through diagnostics, transforms and a model to contributions and ROI.
vignette("journeys"): Building journeys from event logs, and what a naive group_by() and paste() gets wrong.
vignette("carryover"): Choosing decay and lag by cross-validation against your KPI rather than by eye.
vignette("tidymodels"): Tuning carryover, saturation and model penalty jointly in one search.
vignette("spine"): Why carryover and credit are the same idea.

The media half

Carryover: adstock_geometric() for the standard geometric kernel, adstock_weibull() and adstock_delayed() when the response peaks after the spend, adstock_filter() for a kernel of your own. adstock_weights() and adstock_weights_weibull() expose the kernels themselves, and adstock_state() carries a filter across a boundary.
Saturation: saturate_hill(), saturate_exponential(), saturate_michaelis_menten(), saturate_power(), and the saturate() dispatcher.
Composing them: media_transform() applies carryover then saturation, in that order, and will not do it backwards quietly.
Choosing parameters: decay_from_half_life(), half_life() and effective_window() translate between half-lives and decay coefficients. tune_carryover() and tune_carryover_joint() select them by cross-validation against an actual KPI; adstock_steady_state() removes start-up bias.

The attribution half

Journeys: build_paths() turns an event log into journeys, handling the seven things that go wrong on the way.
Credit: credit_linear(), credit_first(), credit_last(), credit_position(), credit_time_decay() and credit_custom(), and the data-driven markov_removal(). attribute() runs several at once and attribution_spread() reports how much the answer depends on which you picked.
Diagnostics: path_summary(), path_lengths(), channel_positions(), assisted_conversions(), top_paths() and conversion_lag(), or path_diagnostics() for all of them.
Interop: as_channel_paths() exports to ChannelAttribution's format.

Reporting, diagnostics and tidymodels

diagnose_media() checks whether the data can support a model at all – run it first. contributions(), roi(), marginal_roi(), mroi(), response_curve() and spend_for() turn a fitted model into a deliverable, and block_bootstrap() puts intervals on it. Each result has a plot() method; mm_palette() exposes their colours. step_adstock() and step_saturation() are recipes steps that carry filter state across the train/test boundary, with carryover_decay() and friends as their dials parameters.

Data

mm_weekly is a synthetic weekly panel generated from known parameters, so a workflow can be checked against the truth it is trying to recover. mm_events is a synthetic touchpoint log containing, on purpose, every case that breaks a naive journey pipeline.

Dependencies

The core transforms need only base R, stats, and cli for error messages. data.table is used by the attribution half, where event logs are large. recipes, dials and the rest of tidymodels are optional and registered conditionally, so the transforms work equally well from a Stan or brms workflow that will never touch them.

Author

Maintainer: Justin Chase jchase.msu@gmail.com

Examples

# The media half: spend in, model-ready regressor out
spend <- c(0, 500, 800, 300, 0, 0, 1200, 400)
media_transform(
  spend,
  adstock = list(decay = decay_from_half_life(2)),
  saturation = list(half_max = 300)
)
#> [1] 0.0000000 0.3280272 0.5296832 0.5213606 0.4350985 0.3525949 0.6088682
#> [8] 0.5985976

# The attribution half: event log in, journeys out
data(mm_events)
paths <- build_paths(mm_events, id = "customer_id", channel = "channel",
                     timestamp = "timestamp", conversion = "conversion")
path_summary(paths)
#>   events_in events_out journeys converting_journeys mean_length median_length
#> 1     20799      17306     6137                2741    2.819945             3
#>   max_length direct_events conversion_events conversions_observed
#> 1          9          1597              2741                 2741
#>   conversions_unattributable
#> 1                          0