Abstract: In order to overcome the shortcomings of the current caching strategy which fails to predict the popularity of short-term hot content sets and cold content sets, this paper proposes a Multi-Arm bandit caching algorithm based on Central Behavior（MACB）. MACB narrows the scope of user group preference by using user central behavior, and then learns the popularity of short-term hot content sets and cold content sets effectively throughout exploration and exploitation process. When applied to China Mobile User Data Record（UDR） data set, MACB makes a hit rate lift compared to the contrast caching strategy, and its effectiveness is shown.

Key words: edge caching, central behavior, multi-arm bandit, hit rate

摘要： 针对现有的边缘缓存策略无法有效预测短时热内容集和冷内容集流行度时变规律，而基于探索的多臂算法缺乏有效机制解决探索过程的过量探索问题，提出了基于用户中心访问行为的多臂缓存方法（MACB）。MACB利用用户中心访问上下文缩小群体访问偏好内容集，在此基础上采用多臂算法的探索开发过程，有效学习短时热内容集和冷内容集的内容流行度变化规律。实验采用了中国移动用户记录数据集，并与相关缓存算法进行对比。结果显示MACB在缓存击中率上均高于其他对比缓存方法，表明了MACB缓存方法的有效性和优越性。

关键词: 边缘缓存, 用户中心访问行为, 多臂算法, 缓存击中率