Hamed Khosravi
← All publications

Catching a Moving Subspace: Low-Rank Bandits Beyond Stationarity

Hamed Khosravi, Xiaoming Huo

arXiv preprint arXiv:2605.20269, 2026

Paper Code

the reward direction alarm the subspace moved rounds → two-window spectral statistic
The reward direction drifts slowly and then turns sharply at change points nobody announces. A two-window spectral statistic crosses its alarm level just after each turn, and the estimate is reset rather than averaged across the break.

The problem

A recommender learns a low-rank structure over user preferences from scalar feedback alone. The structure it is learning does not hold still: it drifts, and occasionally turns. An estimator that keeps averaging across such a turn is fitting a structure that no longer exists.

What we do

Probe, identify the current subspace from the lifted scalar rewards, and exploit by projecting onto it — with a two-window spectral detector that resets the estimate when the subspace shifts.

What it buys you

The low-rank structure keeps the effective number of parameters small enough to learn quickly, without the usual price of assuming that structure is fixed for all time.

Cite

@article{khosravi2026catching,
  title   = {Catching a Moving Subspace: Low-Rank Bandits Beyond Stationarity},
  author  = {Khosravi, Hamed and Huo, Xiaoming},
  journal = {arXiv preprint arXiv:2605.20269},
  year    = {2026}
}