Shapley values for cluster importance

• DOI: 10.1007/s10618-022-00896-3
• Publication Date: 2022-12-06T16:50:52Z
• Authors: Andreas Brandsæter; Ingrid K. Glad
• Keywords: Shapley Value; Complement; Feature (linguistics); Affect; Cooperative game theory; Training set; Value (mathematics); Black box

Abstract This paper proposes a novel approach to explain the predictions made by data-driven methods. Since such rely heavily on data used for training, explanations that convey information about how training affects are useful. The quantify different data-clusters of affect prediction. quantification is based Shapley values, concept which originates from coalitional game theory, developed fairly distribute payout among set cooperating players. A player’s value measure contribution. values often feature importance, ie. features extends this cluster letting clusters act as players in where payouts. methodology proposed lets us explore and investigate any black-box model, allowing new aspects reasoning inner workings prediction model be conveyed users. fundamentally existing explanation methods, providing insight would not available otherwise, should complement including importance.