<ul data-eligibleForWebStory="true"><li>Real-world observational data often contain heterogeneous subpopulations that deviate from global patterns, posing challenges for traditional models.</li><li>DynaSubVAE is introduced as a framework that integrates representation learning and adaptive out-of-domain (OOD) detection.</li><li>The model dynamically updates its latent structure to capture new trends in the data, unlike static approaches.</li><li>DynaSubVAE utilizes a novel non-parametric clustering mechanism inspired by Gaussian Mixture Models to identify and model latent subgroups based on embedding similarity.</li><li>Extensive experiments demonstrate that DynaSubVAE performs well in near-OOD and far-OOD detection, especially excelling in class-OOD scenarios.</li><li>The dynamic subgrouping mechanism of DynaSubVAE outperforms traditional clustering methods like GMM and KMeans++ in terms of OOD accuracy and regret precision.</li></ul>

DynaSubVAE: Adaptive Subgrouping for Scalable and Robust OOD Detection

Discover more