An Adaptive Anxiety Intervention Recommender System Using DSW-UCB Contextual Multi-Armed Bandit with Relevance Filtering

Abstract This research presents the design and evaluation of an adaptive recommender system based on a Discounted Sliding-Window Upper Confidence Bound (DSW-UCB) Contextual Multi-Armed Bandit framework with relevance-filtering updates for personalized anxiety intervention. The study addresses the limitations of traditional full-history CMAB models, which exhibit slow adaptation and historical bias in non-stationary user environments. The proposed model integrates bounded-memory learning, exponential discounting, and relevance-weighted feedback updates to prioritize recent and high-quality interactions while mitigating the influence of obsolete or noisy data. The system was evaluated using a controlled simulation environment under both single and repeated abrupt contextual shifts. Comparative experiments were conducted against multiple baselines. Results show that the proposed DSW-UCB with relevance filtering achieved up to a 20% increase in cumulative reward and approximately 61% reduction in dynamic regret compared to the full-history baseline.