Search

Abstract Clustering is widely used in cloud computing studies to extract vital information. These studies have ignored investigating the potential improvements in clustering quality from better selection of its dimensions and methods. Consequently, developing an automated technique to perform such a selection was not addressed thoroughly. Most of the recent attempts either relied on feature reduction or general non-automated techniques, which were deemed unreliable for sufficient selection. Therefore, we first conducted a comprehensive investigation to study the impact of selecting better clustering dimensions and methods. Our results indicate achieving significant improvement by 15–70% points through better selection. Then, we developed a novel technique (EFection) to detect the best selection in advance using a combination of internal validation metrics (Davies–Bouldin) and the Pearson correlation coefficient. We evaluate our technique’s accuracy by comparing the clustering quality of its suggestions with that of the optimal selection. We then compare EFection’s performance with recent attempts to measure its superiority. Finally, we validate its applicability when adopted in cloud clustering-based studies. The results show that EFection offers high accuracy, around 83%, and surpasses prior art by 11%.

One challenge of studying cloud workload traces is the lack of available users' identities. Therefore, clustering methods were used to address this challenge through extracting these identities from workload traces. For better extraction, it is beneficial to select attributes (columns in the traces) for clustering by using feature selection methods. However, the use of general selection methods requires details that are not available for workload traces (e.g. predefined number of clusters). Therefore, in this paper, we present an unsupervised feature selection method for cloud workload traces to identify good candidate attributes for clustering. This method uses Silhouette coefficients to rank attributes that are best for users' extraction through clustering. The performance of our SeQual method is evaluated in comparison with commonly used (supervised and unsupervised) feature selection methods with the help of clustering quality metrics (i.e. adjusted rand index, entropy and precision). The results show that the SeQual method can compete with the supervised methods and perform better than unsupervised ones, with an average accuracy between 90% and 99%. [ABSTRACT FROM AUTHOR]