A Hierarchical LinUCB Contextual-Bandit Architecture for Adaptive Content Personalization in Nostalgic Persuasive Systems
Year: 2026
Type: article
Abstract: Adaptive persuasive systems increasingly rely on personalized content selection to sustain user engagement and improve intervention effectiveness, but designing such personalization mechanisms presents a fundamental engineering challenge. The instance when a new user first interacts with a system, there is little or no historical information from which to infer individual preferences, yet the system must still make a meaningful recommendation, and as interaction data accumulates, the personalization strategy must also evolve to reflect each user's demonstrated preferences rather than relying solely on population-level behaviour. Balancing these competing requirements constitutes the well known cold-start problem, which remains a persistent challenge in intelligent recommendation and persuasive technologies. Within adaptive persuasive systems, contextual bandits provide an effective mechanism for adaptive decision making by continuously balancing exploration of uncertain recommendation with exploitation of previously successful choices using both contextual information and user feedback, allowing personalization to improve as interactions accumulate. This research presents the design of a hierarchical contextual-bandit layer within a nostalgic persuasive system that, rather than replacing the upstream recommendation models, serves as a final adaptive decision layer selecting personalized nostalgic content from their candidate outputs. The proposed design combines a global Linear Upper Confidence Bound (LinUCB) model that provides robust recommendations during the cold-start phase with progressively introduced per-user LinUCB models enabling increasingly personalized adaptation as interaction history accumulates. This study makes three contributions: First, a hierarchical contextual-bandit architecture integrating global and per-user LinUCB models to address both cold-start initialization and progressive personalization; Second, an efficient context-vector and reward formulation combining real-time affective signals, demographic information, and historical feedback. and Third, a practical deployment blueprint for integrating contextual-bandit learning into emotion-aware persuasive systems, supported by system level field validation reported in a companion study.
Institution: Rivers State University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7204171038