Democratizing Marketing Mix Models (MMM) with Open Source and Gen AI

been within the trade for a number of years and lately they’ve skilled a renaissance. With digitally tracked indicators being deprecated for rising information privateness restrictions, Entrepreneurs are turning again to MMMs for strategic, dependable, privacy-safe measurement and attribution framework.

In contrast to user-level monitoring instruments, MMM makes use of aggregated time-series and cross-sectional information to estimate how advertising and marketing channels drive enterprise KPIs. Advances in Bayesian modeling with enhanced computing energy has pushed MMM again into the middle of promoting analytics.

For years, advertisers and media businesses have used and relied on Bayesian MMM for understanding advertising and marketing channel contributions and advertising and marketing price range allocation.

The Position of GenAI in Trendy MMM

An rising variety of firms are actually using GenAI options as an enhancement to MMM in a number of methods.

1. Knowledge Preparation and Characteristic Engineering
2. Pipeline Automation: Producing code for MMM pipeline
3. Perception Rationalization – translate mannequin insights into plain enterprise language
4. State of affairs planning and price range optimization

Whereas these capabilities are highly effective, they depend on proprietary MMM engines.

The aim of this text is to not showcase how Bayesian MMM works however to show a possible open-source and free system design that entrepreneurs can discover with out the necessity of subscribing to black field MMM stack that distributors within the trade present.

The strategy combines:

1. Google Meridian because the open-source Bayesian MMM engine
2. Open-source Giant Language Mannequin (LLMs) – Mistral 7B as an perception and interplay layer on prime of Meridian’s Bayesian inference output.

Right here is an structure diagram that represents the proposed open-source system design for entrepreneurs.

This open-source workflow has a number of advantages:

1. Democratization of Bayesian MMM: eliminates the black field downside of proprietary MMM instruments.
2. Value Effectivity: reduces monetary barrier for small/medium companies to entry superior analytics.
3. This seperation preserves statistcal rigor required from MMM engines and makes it simply extra accessible.
4. With a GenAI insights layer, audiences don’t want to know the Bayesian math, as a substitute they will simply work together utilizing GenAI prompts to study mannequin insights on channel contribution, ROI, and potential price range allocation methods.
5. Adaptability to newer open-source instruments: a GenAI layer could be changed with newer LLMs as and when they’re brazenly obtainable to get enhanced insights.

Palms-on instance of implementing Google Meridian MMM mannequin with a LLM layer

For the aim of this showcase, I’ve used the open-source mannequin Mistral 7B, sourced regionally from the Hugging Face platform hosted by the Llama engine.

This framework is meant to be domain-agnostic, i.e. any different open-source MMM fashions akin to Meta’s Robyn, PyMC, and many others. and LLM variations for GPT and Llama fashions can be utilized, relying on the size and scope of the insights desired.

Necessary notice:

1. An artificial advertising and marketing dataset was created, having a KPI akin to ‘Conversions’ and advertising and marketing channels akin to TV, Search, Paid Social, E-mail, and OOH (Out-of-Dwelling media).
2. Google Meridian produces wealthy outputs akin to ROI, channel coefficients and contributions in driving KPI, response curves, and many others. Whereas these output are statistically sound, they typically require specialised experience to interpret. That is the place an LLM turns into worthwhile and can be utilized as an perception translator.
3. Google Meridian python code examples had been used to run the Meridian MMM mannequin on the artificial advertising and marketing information created. For extra info on the way to run Meridian code, please check with this page.
4. An open-source LLM mannequin, Mistral 7B, was utilized resulting from its compatibility with the free tier of Google Colab GPU sources and likewise for being an sufficient mannequin for producing instruction-based insights with out counting on any API entry necessities.

Instance: the beneath snippet of Python code was executed within the Google Colab platform:

# Set up meridian: from PyPI @ newest launch 
!pip set up --upgrade google-meridian[colab,and-cuda,schema] 

# Set up dependencies 
import IPython from meridian 
import constants from meridian.evaluation 
import analyzer from meridian.evaluation 
import optimizer from meridian.evaluation 
import summarizer from meridian.evaluation 
import visualizer from meridian.evaluation.overview 
import reviewer from meridian.information 
import data_frame_input_data_builder 
from meridian.mannequin import mannequin
from meridian.mannequin import prior_distribution 
from meridian.mannequin import spec 
from schema.serde import meridian_serde 
import numpy as np 
import pandas as pd

An artificial advertising and marketing dataset (not proven on this code) was created, and as a part of the Meridian workflow requirement, an enter information builder occasion is created as proven beneath:

builder = data_frame_input_data_builder.DataFrameInputDataBuilder( 
   kpi_type='non_revenue', 
   default_kpi_column='conversions', 
   default_revenue_per_kpi_column='revenue_per_conversion', 
   ) 

builder = ( 
   builder.with_kpi(df) 
  .with_revenue_per_kpi(df) 
  .with_population(df) 
  .with_controls( 
  df, control_cols=["sentiment_score_control", "competitor_sales_control"] ) 
  ) 

channels = ["tv","paid_search","paid_social","email","ooh"] 

builder = builder.with_media( 
  df, 
  media_cols=[f"{channel}_impression" for channel in channels], 
  media_spend_cols=[f"{channel}_spend" for channel in channels], 
  media_channels=channels, 
  ) 

information = builder.construct() #Construct the enter information

Configure and execute the Meridian MMM mannequin:

# Initializing the Meridian class by passing loaded information and customised mannequin specification. One benefit of utilizing Meridian MMM is the power to set modeling priors for every channel which supplies modelers capacity to set channel distribution as per historic information of media conduct.

roi_mu = 0.2  # Mu for ROI prior for every media channel.
roi_sigma = 0.9  # Sigma for ROI prior for every media channel.

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

model_spec = spec.ModelSpec(prior=prior, enable_aks=True)

mmm = mannequin.Meridian(input_data=information, model_spec=model_spec)


mmm.sample_prior(500)
mmm.sample_posterior(
    n_chains=10, n_adapt=2000, n_burnin=500, n_keep=1000, seed=0
)

This code snippet runs the meridian mannequin with outlined priors for every channel on the enter dataset generated. The subsequent step is to evaluate mannequin efficiency. Whereas there are mannequin output parameters akin to R-squared, MAPE, P-Values and many others. that may be assessed, for the aim of this text I’m simply together with a visible evaluation instance:

model_fit = visualizer.ModelFit(mmm)
model_fit.plot_model_fit()

Now that the Meridian MMM mannequin has been executed, we’ve mannequin output parameters for every media channel, akin to ROI, response curves, mannequin coefficients, spend ranges, and many others. We will deliver all this info right into a single enter JSON object that can be utilized straight as an enter to the LLM to generate insights:

import json

# Mix every little thing into one dictionary
genai_input = {
    "roi": roi.to_dict(orient='data'),
    "coefficients": coeffs.to_dict(orient='data'),
    "priors": priors.to_dict(orient='data'),
    "response_curves": response_curves.to_dict(orient='data')
}

# Convert to JSON string for the LLM
genai_input_json = json.dumps(genai_input, indent=2)

Downloading Mistral 7B LLM from the Hugging Face platform regionally and putting in the required Llama engine to execute the LLM:

# Obtain the Mistral 7B llm from Hugging Face
!wget -O /content material/fashions/mistral-7b-instruct-v0.2.Q4_K_M.gguf 
https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/resolve/primary/mistral-7b-instruct-v0.2.Q4_K_M.gguf

# Set up Llama for python engine
!pip set up llama-cpp-python --upgrade

Executing the Mistral LLM utilizing the enter JSON having Meridian MMM output and together with the suitable educational immediate:

from llama_cpp import Llama

# Initialize the mannequin
llm = Llama(
    model_path="/content material/fashions/mistral-7b-instruct-v0.2.Q4_K_M.gguf",
    n_ctx=2048,        
    n_gpu_layers=35,    
    seed=42,
    verbose=False,
    temperature=0.2
)

# Convert JSON to a immediate for suggestions
immediate = f"""
You're a advertising and marketing combine modeling skilled.

1. Summarize the effeciency of every advertising and marketing channel primarily based on MMM output
2. Clarify the important thing drivers of ROI for every channel in easy enterprise phrases.
3. Channels with diminishing returns
4. If we wish to maximize whole ROI with the present price range, how ought to we reallocate spend throughout channels

JSON:
{genai_input_json}
"""

# Generate suggestions
response = llm(immediate, max_tokens=300)

# Print solely the textual content output
print(response["choices"][0]["text"])

Instance Output:

1. Based mostly on the Advertising Combine Modeling (MMM) output, TV, OOH, and Search Adverts have a optimistic ROI (3.0, 2.5, and 1.5 respectively), whereas Social Media and E-mail have a decrease ROI (0.9 and 0.7 respectively).

2. The important thing drivers of ROI for every channel in easy enterprise phrases are:

   – TV: Each greenback spent on TV advertisements generates $3 in gross sales.

   – Social Media: Each greenback spent on Social Media advertisements generates $0.9 in gross sales.

   – Search Adverts: Each greenback spent on Search Adverts generates $2.5 in gross sales.

   – E-mail: Each greenback spent on E-mail advertising and marketing generates $0.7 in gross sales.

   – OOH: Each greenback spent on Out of Dwelling (OOH) advertisements generates $1 .5 in gross sales.

3. Channels with diminishing returns are Social Media and E-mail as their response curves present that the gross sales progress will not be proportional to the rising advertising and marketing spend.

4. To maximise whole ROI with the present price range, we must always contemplate reallocating spend from Social Media and E-mail in the direction of TV and OOH as they’ve a better ROI.

Sensible Issues

• Mannequin high quality and insights are nonetheless depending on enter information high quality.
• Immediate design is essential to keep away from deceptive insights.
• Automation for enter information processing and mannequin output reporting and visualization will assist this stack to function at scale.

Closing ideas

This walkthrough illustrates how a possible open-source primarily based Bayesian MMM augmented with a GenAI workflow can translate advanced Bayesian outcomes into actionable insights for entrepreneurs and leaders.

This strategy doesn’t try and simplify the maths behing Advertising Combine Fashions, as a substitute it preserves it and makes an try and make it extra accessible for broader audiences with restricted mannequin information and price range sources for his or her group.

As privacy-safe advertising and marketing analytics turns into a norm, open-source MMM methods with GenAI augmentation supply a sustainable path: clear, adaptable, and designed to evolve with each enterprise and underlying know-how.

Assets & References: