This paper proposes the Selective Reinitialization (SeRe) framework to address the "plasticity loss" problem of the Clustering of Neural Bandits (CNB) algorithm, a neural network-based extension of the clustering technique (CB) of the bandit algorithm. While CNB improves performance by clustering similar bandits, its fixed neural network parameters struggle to adapt to abnormal environments over time. SeRe mitigates plasticity loss and achieves stable knowledge retention by selectively reinitializing underutilized units using a contribution utility metric. Furthermore, it ensures effective adaptation without unnecessary reinitialization through an adaptive change detection mechanism that adjusts the reinitialization frequency based on the degree of abnormality. Theoretically, SeRe achieves sublinear cumulative regret in interval-normal environments. Experiments on six real-world recommendation datasets demonstrate lower regret, improved adaptability, and robustness compared to the existing CNB algorithm.
