Brief

Researchers have introduced a new framework called Deep Spurious Regression (DSR) to address the issue of spurious correlations in continuous prediction tasks. Unlike previous work focused on classification, DSR tackles attribute-label confounding in regression, aiming for models that generalize reliably even when training data exhibits misleading correlations. The proposed method leverages similarities among spurious attributes in both feature and label spaces to improve calibration and robustness across different attribute-label combinations, showing strong performance in computer vision, environmental sensing, and LLM regression. AI