Building Causal AI Models: A Practical Workflow with Robert Osazuwa Ness

Understanding Causal AI's Advantage

• 🎯 Causal AI offers a clear advantage over traditional machine learning when the goal is to understand if one variable directly causes another, rather than just identifying correlations.
• 🔑 This is crucial for problems where understanding the mechanism of influence is key, moving beyond simple predictive relationships.

The Causal AI Workflow

• ❓ The process typically begins with a specific question to answer, such as whether engagement in side quests affects in-game purchases.
• 🗺️ Next, causal assumptions are defined, often visually represented as a Directed Acyclic Graph (DAG), illustrating hypothesized relationships between variables (e.g., guild membership influencing both side quest engagement and purchases).
• 💻 These assumptions are then translated into code using libraries like dowhy, specifying the DAG structure within a model.

Data, Identification, and Estimation

• 🔍 The workflow involves identification, which assesses if the causal question can be answered given the current assumptions and available data, considering potential unobserved confounders.
• 📊 If identification is not possible, strategies include observing additional variables or using techniques like instrumental variable analysis or front-door analysis if mediators are present.
• 📈 Once the question is answerable, the next step is estimation, selecting a statistical approach (e.g., linear regression, propensity scores, double machine learning) to quantify the causal relationship, each with its own statistical trade-offs.

Validation and Sensitivity Analysis

• ✅ Finally, sensitivity analysis is performed to test how robust the conclusions are to potential violations of assumptions, such as missing variables in the DAG, small sample sizes, or strong reliance on linearity.
• 💡 This step helps understand the reliability of the results by examining their sensitivity to the assumptions made throughout the modeling process.