Shapley Chains: Extending Shapley Values to Classifier Chains

In spite of increased attention on explainable machine learning models, explaining multi-output predictions has not yet been extensively addressed. Methods that use Shapley values to attribute feature contributions the decision making are one most popular approaches explain local individual and global predictions. By considering each output separately in tasks, these methods fail provide complete explanations. We propose Chains overcome this issue by including label interdependencies explanation design process. assigns as importance scores classification using classifier chains, separating direct indirect influence scores. Compared existing methods, approach allows a more contribution tasks. mechanism distribute hidden outputs with respect given chaining order outputs. Moreover, we show how our can reveal missed approaches. helps emphasize real factors applications better understanding flow information through synthetic real-world datasets.