Research Area:  Machine Learning
This paper introduces an autonomic method to optimize Feature Selection (FS) in autonomic systems while also presenting a taxonomy of FS techniques. Feature selection is a dimension reduction technique that has been proven to lead to improved performance by avoiding overfitting and to address complexity, thus providing faster and cost-effective algorithms. To be successful, the current FS methods are heavily reliant on two key elements: (1) a well defined and static learning objective, and (2) a relevant dataset. Current FS approaches involve mostly a manual process and do not take into account the constant change in the state of the underlying system. However, the method to automate the FS process presented in this paper, the Autonomic Feature Selection (AFS), uses representation and transfer learning concepts borrowed from the deep learning area. Representations of various types of monitoring data from a system are learned, and the resulting knowledge is transferred and re-used. The AFS is targeted to self-aware software systems, where the learning objective and (or) the availability of the relevant data are subject to change. The results suggest that although building and transferring learned representations is computationally expensive, its benefits outweigh the cost, by offering the potential to build flexible and re-usable knowledge structures.
Author(s) Name:  Patrick Day,Stefano Iannucci,Ioana Banicescu
Journal name:  Future Generation Computer Systems
Publisher name:  Elsevier
Volume Information:  Volume 111, October 2020, Pages 68-81
Paper Link:   https://www.sciencedirect.com/science/article/abs/pii/S0167739X19320564