Energy efficiency has gradually become an important factor for improving the user experience of smartphones. Many background applications continue to consume energy without users' awareness, because these applications use wakelock to prevent the device from going to sleep state, while such misused will cause no-sleep bugs. In this paper, a framework for sensing wakelock application and mining users' preferred application is presented, and the power consumption of smartphone is reduced through shutting down users' non-preferred applications. The central idea is to firstly divide applications into nine categories, then calculate the applications' nine-dimensional vectors cosine similarity with basis vector. The user's preference application types are given based on three parameters: the number of clicks and running time under the foreground and background respectively. The result of vector spaces can directly provide suggestions on user power optimization. In order to verify algorithm's effectiveness, sixty-four people's application behavior data in three weeks are collected. The statistical data show that the proposed framework can efficiently recognize users' preferred applications, and reduce power consumption of smartphone.