An Explainable Multi-Task Similarity Measure: Integrating Accumulated Local Effects and Weighted Fréchet Distance

Explainability enables task comparison: The paper demonstrates that Explainable AI techniques (specifically Accumulated Local Effects curves) can be used to measure similarity between machine learning tasks by examining how features influence predictions, making the comparison process transparent and interpretable rather than a "black box" operation.

Understanding task relationships improves knowledge transfer: By identifying which tasks are genuinely similar through explainable methods, organizations can make better decisions about when to share learning between tasks in multi-task learning scenarios, potentially improving model performance while maintaining clarity about why tasks are being grouped together.

The measure works across different model types: The proposed similarity measure is model-agnostic, meaning it can compare tasks even when different machine learning models are used for each task—an important practical consideration for responsible AI deployment where organizations may use diverse tools and legacy systems.

Feature importance is central to fair comparison: The measure incorporates feature importance weighting and accounts for differences in predictive performance across tasks, addressing a key fairness concern: that tasks shouldn't be considered similar simply because they perform well, but rather because they rely on similar underlying patterns in the data.

Validation spans multiple real-world contexts: Testing on diverse datasets (Parkinson's disease prediction, bike-sharing usage, and facial attribute recognition) suggests the approach has practical applicability across different domains, though organizations should carefully consider domain-specific factors when implementing it.