Exploratory Study of Machine Learning Techniques for Supporting Failure Prediction

Abstract

The growing complexity of software makes it difficult or even impossible to detect all faults before deployment, and such residual faults eventually lead to failures at runtime. Online Failure Prediction (OFP) is a technique that attempts to avoid or mitigate such failures by predicting their occurrence based on the analysis of past data and the current state of a system. Given recent technological developments, Machine Learning (ML) algorithms have shown their ability to adapt and extract knowledge in a variety of complex problems, and thus have been used for OFP. Still, they are highly dependent on the problem at hand, and their performance can be influenced by different factors. The problem with most works using ML for OFP is that they focus only on a small set of prediction algorithms and techniques, although there is no comprehensive study to support their choice. In this paper, we present an exploratory analysis of various ML algorithms and techniques on a dataset containing failure data. The results show that, for the same data, different algorithms and techniques directly influence the prediction performance and thus should be carefully selected.