Zero-Shot Machine Learning Technique for Classification of Multi-User Big Data Workloads

During the last decade machine learning has revolutionized computer science applications. Supervised machine learning algorithms have become especially successful in many industries including health, legal, security, finance, travel, and others. Training supervised learning algorithms, however, is expensive because the real world contains a very large number of different classes that need to be covered by the training set. This is especially true for the highly variable, multi-user workload data produced by the strategically vital big data and cloud software stacks. It is very important, however, to be able to accurately classify these complex workloads in order to enable autonomic management and optimization. Zero-Shot Learning (ZSL) is an advanced machine learning approach that enables classification of objects without having to explicitly train on examples of those objects. In this paper we present a new ZSL technique intended to reduce the expense of assembling workload training sets for big data analytic workloads. We demonstrate that multi-user big data workloads can be treated as hybrids of simpler, single-user workload classes, and classified accurately without having to explicitly train on example instances of multi-user workloads. Our technique is able to accurately classify both unseen multi-user workloads, and seen single-user workloads using the same classifier. We demonstrate 83% classification accuracy for the unseen multi-user workloads, and 92% classification accuracy for the seen, single-user workload classes.