Improving Network Service Fault Prediction Performance with Multi-Instance Learning

Abstract

The internet has undoubtedly become everywhere essential to majority of the people from business services to entertainment. Early prevention of network service faults can greatly improve customer satisfaction and experience. Proactive network service faults prediction can certainly help the internet service providers to reduce service workloads and costs. One of the assumptions often made by standard supervised learning is to treat each session log (generated from the network management system) as an individual instance, where each instance is assigned a class label. Although such assumption is appropriate in some domains, it may not be appropriate in network service fault prediction since a network service fault is represented by a collection of session logs. In this paper, we aim to improve the network service fault prediction by transforming the single-instance learning to a multi-instance learning problem. We evaluate our proposed method on a real-world network data and compared with the baseline single-instance learning method. The multi-instance learning approach achieves a higher AUROC performance to single-instance learning approach.