Service Metric Prediction in Clouds using Transfer Learning

Cloud service management for telecommunication operators is crucial and challengingespecially in a constantly changing operational execution environment. ThusPerformance models can be used to maintain service quality. But building a traditionalperformance model to predict the clients' service quality might require re-training themodels from scratch in the case of environmental changes to maintain predictionaccuracy. And in doing so, a huge data-collection overhead arises that can significantlydegrade the performance, especially if the system needs to perform accuratereal-time predictions. Thus, for the aim of improving the prediction’s accuracy, indynamic environments, we use transfer learning approaches. It re-uses knowledgeobtained from one domain (source domain) to another (target domain). But it isimportant to determine how transferable a source domain is in learning a task in aspecific target domain. In this thesis, we use an information theoretic approach forestimating the performance of transferring representations between domains inclassification problems. We use neural networks models to show a negativecorrelation between the novel metric called H-score, and the prediction’s loss. Thisresults in significant speed up and time efficiency when using the H-score for selectingthe best representations to transfer between domains compared to traditionaltransfer learning approaches. We also find that manual feature selection is a moreviable approach than automatic feature selection especially for transfer learning. Wecollected and evaluated the transfer learning approaches using traces from runningVideo-on-Demand and Key-Value Store services executed on different test-beds