FLEX offers signed feature attributions for multi-label classifiers
A new arXiv paper proposes FLEX, a global surrogate method that explains multi-label classifiers with signed feature attribution lists, claiming high fidelity across benchmarks.
In plain words
Multi-label classifiers assign many labels at once, and a correct label can still rest on shaky reasoning. FLEX is a proposed explanation method that lists, for each label, which input features pushed toward or against it. The authors say it stays accurate while remaining simple to interpret.
What was announced
The paper, posted to arXiv, introduces FLEX (Fisher LDA Explainer), a global surrogate model that fits a single direct multi-label Fisher LDA over the full binary label matrix. This yields signed feature attribution lists for every label within one coherent coordinate system, so each feature's direction of influence is visible per label. The authors report that across the benchmarks and classifier families they evaluated, FLEX reproduces the black box's predictions with consistently high fidelity, and that pooling covariance across labels beats a per-label alternative on most datasets, with the largest gains where labels are numerous and positive examples are scarce.
Beyond fidelity, the authors describe a fidelity-complexity Pareto curve that rises monotonically, meaning practitioners get a predictable path to better explanations as more features are allowed. They add that FLEX's in-sample fidelity matches or leads Binary Relevance alternatives once a modest feature budget is granted. The magnitude of each discriminant weight vector bounds that label's separability, which can flag labels predicted correctly for the wrong reasons. The evaluated classifiers largely agreed on predictions but relied on different feature subsets when no single feature dominated.
What it does not show
- The evidence is the paper's abstract only; results are author-reported on the evaluated benchmarks and have not been independently verified.
- The abstract is truncated mid-sentence, so full details on datasets, baselines and the final findings are not available.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.