Application layer multicast (ALM) which uses end hosts to relay data has a prosperous prospect in one-to-many media content distribution. Compared with IP multicast, ALM is more flexible and deployable. But data delivery in ALM tree can be easily interrupted by departure of end hosts, which may lead to degradation of QoS in time-sensitive applications such as live streaming. Proposed in this paper is a new ALM tree construction algorithm based on the analysis of over 10,000,000 real log traces from a popular live streaming system in China. The algorithm can significantly reduce the interruption times of ALM tree in live streaming and its effectivity has been verified through simulation.