Simple Projection-Free Algorithm for Contextual Recommendation with Logarithmic Regret and Robustness

TL;DR

This paper introduces a simple, efficient, projection-free algorithm for contextual recommendation that achieves logarithmic regret and is robust to suboptimal feedback, improving computational efficiency over prior methods.

Contribution

The paper proposes a novel, simpler algorithm that removes the Mahalanobis projection step, maintaining regret guarantees and robustness in contextual recommendation tasks.

Findings

Achieves $O(d\,\log T)$ regret bound with simpler updates.

Removes Mahalanobis projection, reducing computational complexity.

Maintains robustness to suboptimal action feedback.

Abstract

Contextual recommendation is a variant of contextual linear bandits in which the learner observes an (optimal) action rather than a reward scalar. Recently, Sakaue et al. (2025) developed an efficient Online Newton Step (ONS) approach with an $O(d\log T)$ regret bound, where $d$ is the dimension of the action space and $T$ is the time horizon. In this paper, we present a simple algorithm that is more efficient than the ONS-based method while achieving the same regret guarantee. Our core idea is to exploit the improperness inherent in contextual recommendation, leading to an update rule akin to the second-order perceptron from online classification. This removes the Mahalanobis projection step required by ONS, which is often a major computational bottleneck. More importantly, the same algorithm remains robust to possibly suboptimal action feedback, whereas the prior ONS-based method required running multiple ONS learners with different learning rates for this extension. We describe how our method works in general Hilbert spaces (e.g., via kernelization), where eliminating Mahalanobis projections becomes even more beneficial.