Stochastic Linear Bandits with Partially Observed Actions
arXiv:2607.08971
2026
Architecture
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The transferable contribution is an online low-rank representation procedure for actions whose coordinates are only partially observed. The key engineering asset is the separation between slowly changing representation estimation and within-epoch prediction: estimate a latent subspace from all masked candidates, freeze that subspace for an epoch, reconstruct each new candidate in the subspace, and learn only in the reduced coordinates. This can turn incomplete high-dimensional candidate features into a stable low-dimensional input for a recommender or ranking network, reducing downstream compute and potentially improving sample efficiency when candidate embeddings are strongly low-rank.
Ideas from this paper
Unverified
2026
Add an online low-rank reconstruction layer before a recommender or ranking MLP. It estimates a latent action-feature subspace from masked candidate vectors, freezes that subspace for an epoch, reconstructs each partially observed candidate in the latent coordinates, and feeds only those coordinates to the predictor. The method is most promising when the ambient candidate dimension d is large but the effective rank m is small and missingness is not too severe.
Useful6/10
Difficulty5/10
Novelty7/10