The tutorial focuses on a pilot-scale pit lake system in the Athabasca Oil Sands region, Canda where complex physico-chemical interactions and limited process understanding constrain traditional modeling approaches. In such data-scarce and poorly characterized environments, ML offers a practical alternative, capturing hidden patterns and dependencies to support predictive insight and management decisions. Discussion will be focused on;
- Why existing XAI (e.g., SHAP) can be misleading in correlated multivariate data ?
- How X-VARS bridges local and global interpretations and remains data-size independent ?
- How models with similar prediction accuracy can have different reasoning ?
Finally, the talk will focus on the need for ensemble-based explanation and their reproducibility in ML modeling. This approach strengthens trust and transparency in environmental ML applications, enabling defensible AI decisions for scientists and regulators alike.
Start the conversation