Understanding online content consumption patterns is crucial for a wide range of applications. A challenging problem that has attracted significant interest lately is making ML-based models for content analytics robust to out-of-distribution shifts. Recent work has shown that standard ML model estimation via empirical risk minimization (ERM) can produce models that achieve high accuracy on average but low accuracy on underrepresented groups of users due to the prevalence of spurious features. A predominant approach to tackle this group robustness problem minimizes the worst group error (akin to a minimax strategy) on the training data, hoping it will generalize well on the held-out test data. However, this is often suboptimal since it implicitly assumes high similarity between in-sample and out-of-sample user groups. It also, simplistically, assumes full knowledge of the spurious features that must be controlled for in a given prediction problem. To address these problems, we propose two simple ranking-based approaches that differentially reweight a ranked list of poorly-performing groups in the training data to learn models exhibiting strong OOD performance and do not require a priori knowledge of the spurious features. Empirical evaluation on several datasets highlights the superior generalization ability of our approaches in selecting and learning models robust to group distributional shifts.

This presentation will be held in 2036 Palmer Commons. There will also be a remote viewing option via Zoom.