iSAGE: An Incremental Version of SAGE for Online Explanation on Data Streams

Explainability methods need to adapt to real-world conditions: Traditional AI explanation tools like SAGE only work with static datasets, but many real-world applications (like fraud detection or medical monitoring) operate on continuously arriving data. iSAGE addresses this gap by enabling feature importance explanations in dynamic, streaming environments.

Feature importance explanations can now handle changing data patterns: The method is designed to detect and react to "concept drift"—when the underlying patterns in data change over time. This means stakeholders can understand not just what features matter, but also when their importance changes, which is critical for maintaining trust in AI systems deployed in evolving conditions.

Efficiency matters for practical AI accountability: iSAGE is engineered to be both time- and memory-efficient compared to repeatedly recalculating explanations from scratch. This practical efficiency is essential for deploying explainability in resource-constrained real-world settings, making responsible AI practices more feasible at scale.

Theoretical rigor supports trustworthy explanations: The authors formally prove that iSAGE maintains the same theoretical properties as the original SAGE method, meaning users can rely on the explanations with the same confidence level. This mathematical validation is important for high-stakes applications where explanation reliability directly impacts decisions.

Different explanation approaches serve different purposes: The paper distinguishes between "interventional" explanations (what happens if we remove a feature) and "observational" ones (what correlations exist), giving practitioners flexibility to choose explanations appropriate for their specific accountability needs.