Design of semi automated algorithmic approach for selection of machine learning services in cloud

Abstract

In the global IT market, Cloud Computing is a fully commercialized high-performance computing system that offers everything as a service (XaaS). Recently, the power of commercialization has brought Machine Learning (ML) as a service as MLaaS. Notably, Machine learning services are offered as pay-per-use method by the leading cloud providers such as Amazon, Microsoft and IBM etc. Further, all the ML activities start from data source training to model deployment through single sign in process are best supported by Cloud machine learning service providers (CMLSP). In this distributed environment, cloud broker is operated as a global resource manager that discovers the resource mapping, job submission and monitoring services for the cloud service user (CSU). Hence the performance of the machine learning services gets affected by the performance of the cloud broker considerably. It is a constraint for some of the cloud machine learning service architecture that cloud users are allowed to choose only the predefined listed algorithms for their experimentation. Whereas, Amazon ML services (AMLS) utilize only a single classification or regression algorithm irrespective of the input data source. Consequently, it leads to inadequate trained model with poor output results. On other hand, Microsoft Azure machine learning services (MAMLS) has an excellent drag and drop architecture as well as the all four major services like classification, regression, and clustering and anomaly detection are best supported. Among many predefined algorithms, MAMLS assures more customization through python SDK in relation with the data scientist. When CSU tries all theavailable algorithms to build a decidedly suitable model, it increases the cost of the service in this pay-per-use model. This research aims to build an appropriate technique so as to benefit both CSU and CMLSP through enhancing the utilization of resources more beneficially. And this is achieved by experimenting various characteristic data sources with two leading CMLSP and WEKA