How to Define a New Model Structure

The structure of the Bayesian Network in ConversionFlow is defined in the assets/config.yml file. This guide explains how to customize this structure to create your own models.

Understanding the Model Structure

The model is defined by two key components:

Nodes: These represent the stages or touchpoints in your customer journey.
Edges: These represent the causal relationships or transitions between the nodes.

Steps to Define a New Model

1. Name Your Model

First, give your new model a unique name in the model.name field of the config.yml file.

model:
  name: my_new_model_v1

2. Define the Nodes

The nodes are organized into groups that correspond to the stages of the customer journey. You can add, remove, or rename these groups and the nodes within them.

  nodes:
    start:
      - session_start
    engagement:
      - feature_engagement
      - video_view
    conversion:
      - add_to_cart
      - purchase

3. Define the Edges

The edges define the structure of your Bayesian Network. Each edge is a list of two nodes, representing a directed link from the first node (parent) to the second node (child).

  edges:
    - [session_start, feature_engagement]
    - [session_start, video_view]
    - [feature_engagement, add_to_cart]
    - [video_view, add_to_cart]
    - [add_to_cart, purchase]

4. Define the Priors

For each new edge you create, you need to define a prior distribution for its beta coefficient. This is done in the priors.beta_distributions section.

priors:
  beta_distributions:
    session_start:
      feature_engagement: { distribution: HalfCauchy, sigma: 5 }
      video_view: { distribution: HalfCauchy, sigma: 5 }
    feature_engagement:
      add_to_cart: { distribution: Normal, mu: 0, sigma: 2.5 }
    video_view:
      add_to_cart: { distribution: Normal, mu: 0, sigma: 2.5 }
    add_to_cart:
      purchase: { distribution: HalfCauchy, sigma: 10 }

Parent Node Name: The top-level key (e.g., session_start) is the parent node.
Child Node Name: The nested key (e.g., feature_engagement) is the child node.
Distribution: You can specify any PyMC distribution (e.g., Normal, HalfCauchy, StudentT) and its parameters.

5. Run the Pipeline

Once you have saved your new configuration, you can run the pipeline as usual. The BayesianNetworkModel class will automatically build the new model structure based on your definitions.

python src/orchestration/run_pipeline.py --config assets/config.yml --data your_data.db --output output_new_model

By following these steps, you can create custom Bayesian Network models to analyze any conversion funnel.