Meridian Demo

GitHub - google/meridian

simplified demo

• basic functions & usage

• skips EDA

Step 0: Install

Step 1: Load the data

Read data ⟶ DataFrameInputDataBuilder instance ⟶ InputData object

DataFrameInputDataBuilder:

• is: a class

• purpose: input df ⟶ output InputData

Collect and organize your data

• DV

• IV

  • Media (impressions, clicks, spend)

  • Organic media

  • Non-media treatments

  • Control

  • Seasonality (automatic)

Step 2: Configure the model

How should Meridian interpret our data?

apply priors ⟶ initialize Meridian class ⟶ obtain samples

Initialize the Meridian class:

• input data from step 1

• model specs with priors

Get samples of model parameters

• prior distributions

• posterior distributions

prior = prior_distribution.PriorDistribution(
    roi_m=tfp.distributions.LogNormal(roi_mu, roi_sigma, name=constants.ROI_M)
)

model_spec = spec.ModelSpec(
    prior=prior_distribution.PriorDistribution(),
    media_effects_dist='log_normal',
    hill_before_adstock=False,
    max_lag=8,
    unique_sigma_for_each_geo=False,
    media_prior_type='roi',
    roi_calibration_period=None,
    rf_prior_type='roi',
    rf_roi_calibration_period=None,
    organic_media_prior_type='contribution',
    organic_rf_prior_type='contribution',
    non_media_treatments_prior_type='contribution',
    knots=None,
    baseline_geo=None,
    holdout_id=None,
    control_population_scaling_id=None,
    adstock_decay_spec='geometric',
    enable_aks=False,
)

mmm = model.Meridian(input_data=data, model_spec=model_spec)

Step 3: Run post-modeling quality checks

diagnose issues related to model..

• convergence

• specification

• plausibility

Step 4: Run model diagnostics

Run model diagnostics

continues from step 3, using the methods from the visualizer module.

1. convergence: r-hat statistics.

   1. R-hat close to 1.0 ⟶ convergence

   2. R-hat < 1.2 ⟶ approximate convergence

2. model's fit: compare expected v actual

Step 5: Generate model results & two-page output

Step 6: Run budget optimization & generate an optimization report

Budget optimization scenarios

provide budget ⟶ finds optimal channel allocation

• instantiate BudgetOptimizer class ⟶ run the optimize() method