← AI Techniques Radar
Reasoning & planning

Group-robust learning across heterogeneous feature spaces using worst-group excess loss and shared latent anchors

Define each patient (or group) by the feature space they come from, then train with an objective that focuses on the worst-performing group in terms of excess loss over that group's estimated best achievable loss. Use group-specific encoders that map each group's features into a shared latent space, where learnable Gaussian anchors cluster patients by outcome to enable knowledge transfer even when feature spaces are disjoint.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

In settings where different groups have different sets of available features and may differ in how well their features can support prediction, standard training that simply upweights high-loss examples can over-focus on groups whose errors may stem from fundamentally uninformative features. Optimizing worst-group excess loss instead targets groups relative to what they can realistically achieve, and latent anchors support cross-group collaboration.

What changed

Replaced loss-upweighting of the highest-loss samples with a worst-group excess-loss objective, where the loss is measured against an estimate of each group's achievable optimum. Added group-specific encoders into a shared latent space with learnable Gaussian anchors that cluster patients by outcome.

Where it helps

  • Clinical prediction with patients undergoing different test panels
  • Learning when feature sets are heterogeneous across cohorts
  • Cross-cohort transfer under disjoint or weakly overlapping feature spaces
  • Robust modeling across groups with different intrinsic predictive difficulty

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The method relies on having a way to estimate each group's achievable optimum (needed for excess loss) and assumes that group membership (e.g., which feature space a patient uses) is known and meaningfully partitions the data.

Related World Models papers

Sources