Situation and resource-aware adaptation for ubiquitous data stream mining

Recent advances in sensor technologies and mobile/pervasive computing coupled with increased computational capabilities of mobile devices have led to a new era of data stream mining, named Ubiquitous Data Stream Mining (UDM). UDM is the process of performing real-time and intelligent analysis of continuous data streams onboard mobile/embedded devices. UDM has the potential to benefit a wide range of application domains such as mobile healthcare, Intelligent Transportation Systems (ITS), disaster/emergency management and fire rescue applications. Ubiquitous data stream mining enables mobile users to access real-time analysis of data and make critical decisions on the move. A key distinguishing feature of UDM is the ability to intelligently analyse continuous data streams, learn new patterns and perform predictions in real-time on a relatively resource-limited computational device. To cope with the need to perform UDM analysis, past research has shown that adaptation of the data stream processing with respect to changing levels of resources can significantly improve the continuity of stream mining operations in mobile environments. Resource-aware adaptation is performed by adjusting UDM algorithm parameters such that as resources become scarce, the algorithm continues to operate with reduced accuracy rather than causing the device to shut-down. Alternatively, as the availability of computational resources increases, the adaptation process controls algorithm parameters to increase the accuracy of the analysis. However, such a view of adaptation is limited since it is not underpinned by a holistic application-level strategy. An UDM application’s accuracy requirements vary according to the occurring situations in addition to availability of resources. When the current situation warrants for less frequent monitoring/analysis, the algorithm accuracy can be moderately decreased to preserve resources. On the other hand, in critical situations where there is a need for closer monitoring, it is important to increase the accuracy even if the resource availability is scarce. Therefore, in this dissertation, we propose, develop and validate a novel situationaware adaptation framework for UDM, named Situation and Resource-Aware Adaptation (SARA). Our SARA framework makes two important contributions. IV Firstly, SARA integrates situation-awareness into the adaptation process that controls UDM algorithms and introduces two innovative adaptation mechanisms including situation-aware and hybrid strategies to cater for combinations of scenarios that could arise from changes in situations and availability of computational resources. Secondly, to achieve situation-awareness and identify the occurring situation, we propose and develop as part of SARA an innovative context modelling and reasoning approach, termed Fuzzy Situation Inference (FSI), to represent real-world situations as well as uncertainty associated with these situations. FSI is utilised by SARA to enable a gradual, smooth and fine-grained adaptation of UDM algorithms’ settings according to application constraints. The FSI approach captures the delta changes of context/situations as well as the transition of a situation to another. We leverage the FSI approach, and develop novel adaptation strategies for UDM that take into account both an UDM application’s situation and resource levels. Thus we consider both an application’s needs for analysis accuracy as well as computational resource levels on the mobile device, resulting in improved efficiency and usage of appropriate adaptation strategies. We have implemented our SARA framework on a mobile phone and conducted extensive experimental validation of the proposed strategies and approaches. This evaluation clearly demonstrates the ability of the SARA framework to identify and use changing situations of applications and resource levels to improve the overall lifetime/running time of UDM applications. Thus, we establish that a holistic adaptation of UDM algorithms/applications driven by situation-awareness of the application needs combined with resource consideration is an important step towards realising the full potential of this technology. Such an integrated adaptation makes a significant improvement to the overall efficiency and effectiveness of ubiquitous data stream mining. The contributions of this thesis have resulted in one book chapter, journal article, and six international conference papers.