A Comparative Study of Multi-Armed Bandit Algorithms for Dynamic Filter and Effect Recommendations on Tiktok

College of Liberal Arts & Sciences, University of Illinois Urbana-Champaign, Champaign, Illinois, 61820, USA

jz133@illinois.edu

Abstract

This study assesses how three benchmark Multi-Armed Bandit (MAB) methods—ε-greedy, Upper Confidence Bound (UCB), and Thompson sampling—perform in dynamically recommending filters and effects for short-video platforms. An experiment was configured using real user engagement metrics (likes, shares, views), enabling each algorithm to incrementally build reward estimates and adapt to evolving content trends. The study examines the dual impact of striking a diversity–satisfaction equilibrium on recommendation effect identification. Results suggest that moderate exploration strategies yield the highest effectiveness by enabling the discovery of emerging, underused effects without compromising user satisfaction. A hybrid MAB model is then proposed, in which exploration adjusters are reactive to performance metrics, allowing prompt responses to content trends while maintaining recommendation quality. Finally, practical aspects of deploying these models—such as latency, scalability, and fairness—are discussed, along with concrete recommendations for designing adaptive social media recommendation systems. By integrating conceptual and practical approaches, this study demonstrates how MAB algorithms can achieve both content discovery and reliability.