Clarity in complexity: how aggregating explanations resolves the disagreement problem

• Published in: The Artificial intelligence review Vol. 57; no. 12; p. 338
• Main Authors: Mitruț, Oana, Moise, Gabriela, Moldoveanu, Alin, Moldoveanu, Florica, Leordeanu, Marius, Petrescu, Livia
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 19.10.2024; Springer Nature B.V
• Subjects: Algorithms; Alignment; Artificial Intelligence; Citrus fruits; Classification; Computer Science; Datasets; Machine learning; Case based reasoning; Explainability; Disagreement problem; Rashômon effect; Classification
• ISSN: 1573-7462; 0269-2821; 1573-7462
• DOI: 10.1007/s10462-024-10952-7

The Rashômon Effect, applied in Explainable Machine Learning, refers to the disagreement between the explanations provided by various attribution explainers and to the dissimilarity across multiple explanations generated by a particular explainer for a single instance from the dataset (differences between feature importances and their associated signs and ranks), an undesirable outcome especially in sensitive domains such as healthcare or finance. We propose a method inspired from textual-case based reasoning for aligning explanations from various explainers in order to resolve the disagreement and dissimilarity problems. We iteratively generated a number of 100 explanations for each instance from six popular datasets, using three prevalent feature attribution explainers: LIME, Anchors and SHAP (with the variations Tree SHAP and Kernel SHAP) and consequently applied a global cluster-based aggregation strategy that quantifies alignment and reveals similarities and associations between explanations. We evaluated our method by weighting the -NN algorithm with agreed feature overlap explanation weights and compared it to a non-weighted -NN predictor, having as task binary classification. Also, we compared the results of the weighted -NN algorithm using aggregated feature overlap explanation weights to the weighted -NN algorithm using weights produced by a single explanation method (either LIME, SHAP or Anchors). Our global alignment method benefited the most from a hybridization with feature importance scores (information gain), that was essential for acquiring a more accurate estimate of disagreement, for enabling explainers to reach a consensus across multiple explanations and for supporting effective model learning through improved classification performance.