Alexander Galozy

Alexander Galozy, Slawomir Nowaczyk

In the latent bandit problem, the learner has access to reward distributions and -- for the non-stationary variant -- transition models of the environment. The reward distributions are conditioned on the arm and unknown latent states. The goal is to use the reward history to identify the latent state, allowing for the optimal choice of arms in the future. The latent bandit setting lends itself to many practical applications, such as recommender and decision support systems, where rich data allows the offline estimation of environment models with online learning remaining a critical component. Previous solutions in this setting always choose the highest reward arm according to the agent's beliefs about the state, not explicitly considering the value of information-gathering arms. Such information-gathering arms do not necessarily provide the highest reward, thus may never be chosen by an agent that chooses the highest reward arms at all times. In this paper, we present a method for information-gathering in latent bandits. Given particular reward structures and transition matrices, we show that choosing the best arm given the agent's beliefs about the states incurs higher regret. Furthermore, we show that by choosing arms carefully, we obtain an improved estimation of the state distribution, and thus lower the cumulative regret through better arm choices in the future. We evaluate our method on both synthetic and real-world data sets, showing significant improvement in regret over state-of-the-art methods.

Alexander Galozy, Sadi Alawadi, Victor Kebande et al.

This paper investigates the issue of privacy in a learning scenario where users share knowledge for a recommendation task. Our study contributes to the growing body of research on privacy-preserving machine learning and underscores the need for tailored privacy techniques that address specific attack patterns rather than relying on one-size-fits-all solutions. We use the latent bandit setting to evaluate the trade-off between privacy and recommender performance by employing various aggregation strategies, such as averaging, nearest neighbor, and clustering combined with noise injection. More specifically, we simulate a linkage attack scenario leveraging publicly available auxiliary information acquired by the adversary. Our results on three open real-world datasets reveal that adding noise using the Laplace mechanism to an individual user's data record is a poor choice. It provides the highest regret for any noise level, relative to de-anonymization probability and the ADS metric. Instead, one should combine noise with appropriate aggregation strategies. For example, using averages from clusters of different sizes provides flexibility not achievable by varying the amount of noise alone. Generally, no single aggregation strategy can consistently achieve the optimum regret for a given desired level of privacy.

Alexander Galozy

Reinforcement learning (RL) agents under partial observability often condition actions on internally accumulated information such as memory or inferred latent context. We formalise such information-conditioned interaction patterns as behavioural dependency: variation in action selection with respect to internal information under fixed observations. This induces a probe-relative notion of $ε$-behavioural equivalence and a within-policy behavioural distance that quantifies probe sensitivity. We establish three structural results. First, the set of policies exhibiting non-trivial behavioural dependency is not closed under convex aggregation. Second, behavioural distance contracts under convex combination. Third, we prove a sufficient local condition under which gradient ascent on a skewed mixture objective decreases behavioural distance when a dominant-mode gradient aligns with the direction of steepest contraction. Minimal bandit and partially observable gridworld experiments provide controlled witnesses of these mechanisms. In the examined settings, behavioural distance decreases under convex aggregation and under continued optimisation with skewed latent priors, and in these experiments it precedes degradation under latent prior shift. These results identify structural conditions under which probe-conditioned behavioural separation is not preserved under common policy transformations.

Alexander Galozy, Slawomir Nowaczyk, Mattias Ohlsson

We propose a new sequential decision-making setting, combining key aspects of two established online learning problems with bandit feedback. The optimal action to play at any given moment is contingent on an underlying changing state which is not directly observable by the agent. Each state is associated with a context distribution, possibly corrupted, allowing the agent to identify the state. Furthermore, states evolve in a Markovian fashion, providing useful information to estimate the current state via state history. In the proposed problem setting, we tackle the challenge of deciding on which of the two sources of information the agent should base its arm selection. We present an algorithm that uses a referee to dynamically combine the policies of a contextual bandit and a multi-armed bandit. We capture the time-correlation of states through iteratively learning the action-reward transition model, allowing for efficient exploration of actions. Our setting is motivated by adaptive mobile health (mHealth) interventions. Users transition through different, time-correlated, but only partially observable internal states, determining their current needs. The side information associated with each internal state might not always be reliable, and standard approaches solely rely on the context risk of incurring high regret. Similarly, some users might exhibit weaker correlations between subsequent states, leading to approaches that solely rely on state transitions risking the same. We analyze our setting and algorithm in terms of regret lower bound and upper bounds and evaluate our method on simulated medication adherence intervention data and several real-world data sets, showing improved empirical performance compared to several popular algorithms.