Skip to content

Provably Efficient Multi-Objective Bandit Algorithms under Preference-Centric Customization

Conference: AAAI 2026 arXiv: 2502.13457 Code: None Area: Online Learning / Multi-Armed Bandits Keywords: multi-objective bandits, user preference, Pareto optimality, UCB, regret bound, preference estimation

TL;DR

This paper presents the first theoretical study of preference-aware customization in multi-objective multi-armed bandits (MO-MAB) with explicit user preferences. It proposes the PAMO-MAB framework and designs PRUCB-UP and PRUCB-HP algorithms for the "unknown preference" and "hidden preference" settings, respectively. Through a two-component architecture combining preference estimation and preference-aware optimization, both algorithms achieve near-optimal regret bounds. The paper also proves that preference-free algorithms inevitably incur \(\Omega(T)\) linear regret when the Pareto front contains conflicting arms.

Background & Motivation

Core Problem

Classical MO-MAB formulations target Pareto optimality and employ a global policy to select arms for all users. However, different users hold distinct preferences over objective dimensions:

  • Illustrative scenario: In restaurant recommendation, User 1 prioritizes taste while User 2 prioritizes price; the same restaurant on the Pareto front may satisfy one user but not the other.
  • Key Challenge: When the Pareto front contains two or more conflicting arms (\(|\mathcal{O}^*| \geq 2\)), any algorithm that ignores user preferences will necessarily incur linear regret for some users.

Limitations of Prior Work

  1. Pareto front estimation methods (Pareto-UCB, Pareto-TS): Estimate the full front and select arms uniformly, without user-level customization.
  2. Scalarization methods (S-UCB, GGI): Compress multi-dimensional rewards into a scalar via a fixed scalarization function that is independent of individual users.
  3. Lexicographic MO-MAB: Imposes a strict hierarchy over objectives and cannot model continuous preference trade-offs across users.

Paper Goals

  • Formalize the Preference-Aware MO-MAB (PAMO-MAB) problem, introducing the first explicit user preference framework.
  • Prove an \(\Omega(T)\) lower bound for preference-free algorithms (Proposition 1).
  • Propose near-optimal algorithms for both the hidden-preference and unknown-preference settings.

Method

Problem Formulation

Consider \(N\) users, \(K\) arms, and \(D\) objective dimensions. At each round \(t\), user \(n\) has a stochastic preference \(\boldsymbol{c}_t^n \in \mathbb{R}^D\) with mean \(\bar{\boldsymbol{c}}^n\). Upon selecting arm \(a_t^n\), the agent observes a \(D\)-dimensional reward \(\boldsymbol{r}_{a_t^n, t}\). The overall reward is defined as the inner product \(g_{a_t^n, t} = {\boldsymbol{c}_t^n}^\top \boldsymbol{r}_{a_t^n, t}\), and the objective is to minimize cumulative regret:

\[R(T) = \sum_{t=1}^T \sum_{n=1}^N \bar{\boldsymbol{c}}^{n\top}(\boldsymbol{\mu}_{a_t^{n*}} - \boldsymbol{\mu}_{a_t^n})\]

Overall Architecture: Preference Estimation + Preference-Aware Optimization

All algorithms share a unified two-component framework:

  1. Preference Estimation: Infer user preference vectors from bandit feedback.
  2. Preference-Aware Optimization: Select arms based on preference estimates to align decisions with user intent.

Setting 1: Hidden Preference (PRUCB-HP)

Users provide per-objective ratings and an overall rating; preferences must be inferred from their relationship. Two unique challenges arise:

Challenge 1: Stochastic Mapping Problem

The overall reward satisfies \(g_t = (\bar{\boldsymbol{c}} + \boldsymbol{\zeta}_t)^\top \boldsymbol{r}_t\), where the residual noise \(\boldsymbol{\zeta}_t^\top \boldsymbol{r}_t\) depends on the input (larger rewards yield larger noise), invalidating standard regression models.

Challenge 2: Local vs. Global Exploration Dilemma

  • Local exploration: Select high-reward arms to identify better arms.
  • Global exploration: Select diverse arms to learn preferences.
  • These two objectives conflict: high-reward arms are actually detrimental to preference estimation due to larger noise.

Key Design I: WLS Preference Estimator

Weighted Least Squares (WLS) is used for preference learning, with weights set to the reciprocal of the squared \(\ell_2\) norm of the reward:

\[w_t^n = \frac{\omega}{\|\boldsymbol{r}_{a_t^n, t}\|_2^2}\]
  • High-reward samples receive small weights → suppresses large residual noise.
  • Low-reward samples receive large weights → amplifies their contribution to estimation.
  • Core Lemma: After the re-weighting transformation, the residual noise becomes input-independent \(R'\)-sub-Gaussian with \(R' = \sqrt{\omega}R\), restoring standard regression conditions.

Key Design II: Dual Exploration Strategy

The arm selection policy combines two bonus terms:

\[a_t = \arg\max_{i \in \mathcal{A}_t^n} (\hat{\boldsymbol{c}}_t^n)^\top \hat{\boldsymbol{r}}_{i,t} + B_{i,t}^{n,r} + B_{i,t}^{n,c}\]
  • Reward bonus \(B^{n,r}\): Standard UCB bonus based on Hoeffding's inequality, encouraging local exploration.
  • Preference bonus \(B^{n,c}\): Bonus based on pseudo information gain, \(\beta_t \|\hat{\boldsymbol{r}}_{i,t} + \rho_{i,t}^\alpha \boldsymbol{e}\|_{(\boldsymbol{V}_{t-1}^n)^{-1}}\), encouraging global exploration.

Setting 2: Unknown Preference (PRUCB-UP)

Users explicitly provide preference feedback after arm selection. This setting is considerably simpler than the hidden preference case:

  • Preference estimation: Directly uses the empirical mean \(\hat{\boldsymbol{c}}_{t+1}^n = ((t-1)\hat{\boldsymbol{c}}_t^n + \boldsymbol{c}_t^n) / t\).
  • Optimization strategy: Only the reward bonus is needed; no preference bonus is required: \(a_t^n = \arg\max_{i} (\hat{\boldsymbol{c}}_t^n)^\top (\hat{\boldsymbol{r}}_{i,t} + \rho_{i,t}^\alpha \boldsymbol{e})\).

Theoretical Analysis

Lower Bound (Proposition 1)

When \(|\mathcal{O}^*| \geq 2\), any preference-free algorithm incurs \(R(T) = \Omega(T)\) for some subset of users. The proof argues that any preference-free algorithm achieving sublinear regret under preference \(\boldsymbol{c}_1\) must use the same policy under the opposing preference \(\boldsymbol{c}_2\); since the optimal arm changes but the policy does not, linear regret is unavoidable.

Upper Bound for PRUCB-HP (Theorem 1)

In the hidden preference setting, for \(t \geq M\) (where \(M\) is a constant independent of \(T\)), the asymptotic regret satisfies:

\[R(T) = \tilde{O}(D\sqrt{T})\]

The key technical challenge is bounding the cumulative preference bonus \(\sum_t B_{i,t}^c\), which requires handling the inconsistency between the Gram matrix \(\boldsymbol{V}_t\) constructed from observed rewards and the estimated rewards used in the bonus. This is resolved by transitioning to the expected Gram matrix \(\mathbb{E}[\boldsymbol{V}_t]\).

Upper Bound for PRUCB-UP (Theorem 2)

In the unknown preference setting:

\[R(T) = O(KN\delta \log T + KN\delta D^2)\]

The regret attributable to preference estimation error constitutes only a constant term in \(D\) and \(\delta\), indicating that preference learning has a limited impact on total regret.

Key Experimental Results

Main Results: Regret Comparison Across Preference Settings

Algorithm Preference Setting Regret Trend Sublinear? Key Feature
PRUCB-HP Hidden preference Sublinear growth WLS preference estimation + dual exploration
PRUCB-UP Unknown preference Sublinear growth Empirical mean preference estimation
PRUCB-UP (GT) Known preference Sublinear growth Ground-truth preference upper bound
S-UCB Both settings Linear growth Fixed equal-weight scalarization
S-MOSS Both settings Linear growth Fixed equal-weight + MOSS policy
Pareto-UCB Both settings Linear growth Estimates Pareto front
Pareto-TS Both settings Linear growth Thompson sampling
OFUL Hidden preference Linear growth Linear bandit baseline
UCB / MOSS Hidden preference Linear growth Uses only scalar feedback

Ablation Study: WLS Preference Estimator vs. Standard Linear Regression

Sample Size WLS Estimator Error Standard LR Error WLS Advantage
20 Lower Higher Advantage visible even with few samples
50 Clearly below LR Moderate bias Gap widens
80 Approaches true value [0.5, 0.5] Bias remains WLS converges faster
200 Nearly unbiased Bias persists WLS consistently superior

Experimental setup: 2D toy instance; Arm-1 mean [0.2, 0.2] (dominated), Arm-2 mean [0.8, 0.8] (Pareto-optimal). True preference mean [0.5, 0.5], variance 0.05. \(T=5000\), averaged over 10 repetitions.

Highlights & Insights

  1. Significant theoretical contribution: The paper provides the first proof that preference-free algorithms necessarily fail (\(\Omega(T)\) lower bound) when the Pareto front contains conflicting arms, establishing a rigorous theoretical motivation for preference-aware MO-MAB customization.
  2. Counter-intuitive finding: Pareto-optimal (high-reward) arms are actually detrimental to preference learning — larger rewards amplify the noise introduced by stochastic preferences, overturning the intuition that "better arms carry more information."
  3. Elegant WLS weight design: Setting \(w_t = \omega / \|\boldsymbol{r}_t\|_2^2\) eliminates the input dependence of residual noise, transforming a non-standard regression problem into a canonical sub-Gaussian noise model.
  4. Novel dual exploration mechanism: The pseudo information gain bonus can estimate the contribution of selecting a given arm to preference learning before observing rewards, effectively balancing local and global exploration.
  5. Strong framework unification: PRUCB-UP can be viewed as a special case of PRUCB-HP with zero preference noise; both algorithms share the same algorithmic skeleton.

Limitations & Future Work

  1. Synthetic data only: Experiments are conducted exclusively on synthetic data, without validation on real recommendation or rating systems.
  2. Linear preference assumption: The overall reward is defined as an inner product of preference and reward vectors (linear model), which cannot capture nonlinear user satisfaction functions.
  3. Preference independence assumption: Assumption 3 requires preferences to be independent of rewards, which may not hold in certain scenarios (e.g., price-sensitive users may shift preferences in response to high-priced items).
  4. Preference stationarity: Each user's preference mean \(\bar{\boldsymbol{c}}^n\) is assumed fixed; preference drift is not considered.
  5. Scalability: The WLS estimator requires maintaining a \(D \times D\) Gram matrix and its inverse, incurring high computational cost in high-dimensional objective spaces.
  6. User rotation protocol: The block-rotation protocol used in experiments is somewhat specific and may not generalize to more realistic recommendation scenarios.
  • Pareto front estimation MO-MAB: Pareto-UCB (Drugan & Nowe 2013), Knowledge Gradient (Yahyaa et al. 2014), Thompson Sampling (Yahyaa & Manderick 2015), multi-objective contextual bandits (Turgay et al. 2018; Lu et al. 2019).
  • Scalarization MO-MAB: GGI optimization (Busa-Fekete et al. 2017), multi-objective optimization for music platforms (Mehrotra et al. 2020), Pareto regret analysis (Xu & Klabjan 2023).
  • Lexicographic preference MO-MAB: Hüyük & Tekin 2021 (introduction of lexicographic ordering), Cheng et al. 2024 (hybrid Pareto-lexicographic approach).
  • Linear bandit foundational techniques: OFUL (Abbasi-Yadkori et al. 2011), MOSS (Audibert & Bubeck 2009), variance-dependent bandits (Zhou et al. 2021).

Rating

Dimension Score (1–5) Notes
Novelty 4.5 First theoretical formalization of user-preference-aware MO-MAB customization
Theoretical Depth 5 Complete lower and upper bound analysis with near-optimal regret guarantees
Experimental Thoroughness 3 Synthetic experiments only, but covers diverse settings and ablations
Practical Value 3.5 Promising for recommendation/rating systems; linear preference assumption is a notable limitation
Writing Quality 4 Clear structure, intuitive motivating figures, rigorous theoretical derivations
Overall 4 A pioneering work on preference-centric MO-MAB customization with solid theoretical contributions