Using machine learning for software aging detection in Android system

Software aging is a common experience in Android operating system, as the gradual performance degradation is usually complained by the users. However, the mathematical modelling and detection of such experience is still an emerging issue due to the complexity and relatively young age of Android. This paper applies and compares three machine learning algorithms, namely decision tree, Support Vector Machine (SVM), and Deep Belief Network (DBn), for the detection of software aging in Android. In addition to the traditional aging indicator of launch time (LT), this paper also investigates the effectiveness of page fault (PF) and multiple labels (i.e., the combination of LT and PF). Experimental results show that the accuracy of DBN is comparable to decision tree and SVM when the data volume increases to 5000, which means DBN and other similar algorithms suitable for high dimensional and large data may also play a role in software aging. The results also reveal that PF is a little more stable than LT in terms of variance of accuracy, and it can also be used a good indicator for aging of Android. However, the performance of multiple labels is not improved in our experiments. The observations of this paper are supposed to assist further analysis of software aging in Android and the design of corresponding rejuvenation methods.