Enhancing Context Specifications for Dependable Adaptive Systems: A Data Mining Approach.

Abstract

Context: Adaptive systems are expected to cater for various operational contexts by having multiple strategies in achieving their objectives and the logic for matching strategies to an actual context. The prediction of relevant contexts at design time is paramount for dependability. With the current trend on using data mining to support the requirements engineering process, this task of understanding context for adaptive system at design time can benefit from such techniques as well. Objective: The objective is to provide a method to refine the specification of contextual variables and their relation to strategies for dependability. This refinement shall detect dependencies between such variables, priorities in monitoring them, and decide on their relevance in choosing the right strategy in a decision tree. Method: Our requirements-driven approach adopts the contextual goal modelling structure in addition to the operationalization values of sensed information to map contexts to the system’s behaviour. We propose a design time analysis process using a subset of data mining algorithms to extract a list of relevant contexts and their related variables, tasks, and/or goals. Results: We experimentally evaluated our proposal on a Body Sensor Network system (BSN), simulating 12 resources that could lead to a variability space of 4096 possible context conditions. Our approach was able to elicit subtle contexts that would significantly affect the service provided to assisted patients and relations between contexts, assisting the decision on their need, and priority in monitoring. Conclusion: The use of some data mining techniques can mitigate the lack of precise definition of contexts and their relation to system strategies for dependability. Our method is practical and supportive to traditional requirements specification methods, which typically require intense human intervention.