<ul><li>Machine learning models can face a decrease in performance when faced with data that differs from their training set, known as concept shift.</li><li>Concept shift occurs when there is a change in the distribution of labels conditioned on features, leading to the learning of an incorrect representation by even well-tuned ML models.</li><li>A model called SGShift is proposed for detecting concept shift in tabular data and attributing reduced model performance to shifted features using a Generalized Additive Model (GAM) and feature selection.</li><li>Experiments show that SGShift can identify shifted features with high accuracy, outperforming baseline methods with AUC >0.9 and recall >90%.</li></ul>

Explaining Concept Shift with Interpretable Feature Attribution

Discover more