Document Type

Conference Paper


Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence


Electrical and electronic engineering

Publication Details

2018 IEEE International Conference on Cloud Computing Technology and Science (CloudCom)


We are on the cusp of an era where we can responsively and adaptively predict future network performance from network device statistics in the Cloud. To make this happen, regression-based models have been applied to learn mappings between the kernel metrics of a machine in a service cluster and service quality metrics on a client machine. The path ahead requires the ability to adaptively parametrize learning algorithms for arbitrary problems and to increase computation speed. We consider methods to adaptively parametrize regularization penalties, coupled with methods for compensating for the effects of the time-varying loads present in the system, namely load-adjusted learning. The time-varying nature of networked systems gives rise to the need for faster learning models to manage them; paradoxically, models that have been applied have not explicitly accounted for their time-varying nature. Consequently previous studies have reported that the learning problems were ill-conditioned -the practical, undesirable consequence of this is variability in prediction quality. Subset selection has been proposed as a solution. We highlight the short-comings of subset selection. We demonstrate that load-adjusted learning, using a suitable adaptive regularization function, outperforms current subset selection approaches by 10% and reduces computation.