Reproduce “Global Aggregations of Local Explanations for Black Box models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ilse van der Linden, Hinda Haned, Evangelos Kanoulas
Researchers at the University of Amsterdam developed improved methods for Global Aggregations of Local Explanations (GALE) to synthesize holistic insights from black box models. Their proposed homogeneity-weighted importance significantly outperforms prior methods like LIME's original global importance, demonstrating a more reliable identification of globally influential and class-distinguishing features, particularly in multiclass text classification tasks.
Key insights
- The effectiveness of aggregating local explanations to derive global insights heavily depends on the chosen aggregation function, with simpler methods often being unreliable.
- Existing global importance metrics, such as LIME's original formulation, suffer from biases towards frequently occurring but non-discriminative features, providing limited global understanding.
- Aggregating local explanations, especially by weighting features based on the homogeneity of their influence across classes, can effectively identify truly distinguishing features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.