UniMiner: Towards a unified framework for data mining

Abstract

Wearable devices and Smartphones generate huge data streams in pervasive and ubiquitous environments. Traditionally, big data systems collect all the data at a central data processing system (DPS). These data silos are further analyzed to generate approximated patterns for different application areas. This approach has one-sided utility (i.e. at big data processing end) but two main side-effects that lead towards user's dissatisfaction and extra computational costs. These effects are: (1) since all the data is being collected at central DPS, user privacy is compromised and (2) the collection of huge raw data streams, most of which could be irrelevant, at central systems needs more computational and storage resources hence increases the overall operational cost. Keeping in view these limitations, we are proposing a unified framework that balances between utility and cost of big data system with increased user satisfaction. We studied different data mining systems and proposed a new framework, named as UniMiner, to leverage data mining systems with wearable devices, smartphones, and cloud computing technologies. The gist of UniMiner is the scalability of data mining tasks from resource-constraint devices to collaborative and hybrid execution models. This scalable unified data mining approach distinguishes UniMiner from existing systems by enabling maximum data processing near data sources. Finally, we assessed the feasibility of mobile devices using six frequent pattern mining algorithms. The results show that mobile devices could be adopted as data mining platforms by tuning some additional parameters.