I. Introduction
Device failures happen constantly in modern large-scale storage systems which consist of thousands even millions of devices and disk failures dominate 76-95% of all failed components [1]. The disk failures affect the availability of cloud storage system and hurt the quality of service (QoS). Recently, a host of ML based disk failure detection methods have been proposed to proactively predict forthcoming disk failures so that the operators can guarantee the QoS via timely disk replacement and VM migration.