Why Bayesian Networks?

This document explains why Bayesian Networks were chosen as the core methodology for the ConversionFlow library.

The Challenge: Understanding Complex Customer Journeys

Modern customer journeys are complex and non-linear. A user might interact with a website in many different ways before making a purchase. To optimize this journey, we need to understand the causal relationships between different touchpoints. For example:

• Does watching a video make a user more likely to configure a car?

• Is a user who downloads a brochure more likely to request a test drive?

Answering these questions requires a modeling technique that can handle causality, uncertainty, and complex dependencies.

Why Bayesian Networks are a Good Fit

Bayesian Networks are a type of probabilistic graphical model that are exceptionally well-suited for this task. Here’s why:

1. Modeling Causality

Bayesian Networks are built on the principles of conditional probability and directed acyclic graphs (DAGs). This allows them to explicitly model causal relationships. Each edge in the network represents a direct causal influence from a parent node to a child node.

2. Handling Uncertainty

The real world is full of uncertainty. Bayesian inference allows us to quantify this uncertainty. Instead of giving us a single “best” answer, a Bayesian model provides a full probability distribution for each parameter. This allows us to say not just what we think the effect of a touchpoint is, but also how certain we are about that effect.

3. Incorporating Prior Knowledge

Bayesian models allow us to incorporate prior knowledge into the model through the use of priors. This is a powerful feature that allows us to combine expert domain knowledge with the data. For example, if we have a strong belief that downloading a brochure should have a positive effect on requesting a test drive, we can encode this belief in our choice of prior.

4. Flexibility and Extensibility

Bayesian Networks are highly flexible. The structure of the network can be easily modified to test new hypotheses about the customer journey. Adding or removing nodes and edges is straightforward, allowing for rapid iteration and experimentation.

Comparison to Other Methods

Method	Strengths	Weaknesses
Bayesian Networks	Causal inference, uncertainty quantification	Can be computationally intensive
Logistic Regression	Simple, fast, interpretable	Does not model complex dependencies well
Random Forests	Good predictive power, handles non-linearities	Not easily interpretable, does not model causality
Markov Chains	Good for modeling sequences	Assumes the “Markov property” (memorylessness), which may not hold

While other methods have their strengths, Bayesian Networks provide the best combination of causal inference, uncertainty quantification, and flexibility for the task of modeling and optimizing complex customer journeys.