SELF-CONFIGURING DATA MINING FOR UBIQUITOUS COMPUTING
Electronics Engineering & Computer Science, Ph.D. Dissertation 2013
Assoc. Prof. Yücel Saygın (Thesis Supervisor), Asst. Prof. Gürdal Ertek, Assoc. Prof. Albert Levi, Assoc. Prof. Erkay Savaş, Assoc. Prof. Ernestina Menasalvas
Date &Time: 25 January 2013, 14:30
Place: FENS G032
Keywords: Data Mining, Ubiquitous Computing, Machine learning
Ubiquitous computing software needs to be autonomous so that essential decisions such as how to configure its particular execution are self-determined. Moreover, data mining serves an important role for ubiquitous computing by providing intelligence to several types of ubiquitous computing applications. Thus, automating ubiquitous data mining is also crucial. We focus on the problem of automatically configuring the execution of a ubiquitous data mining algorithm. In our solution, we generate configuration decisions in a resource-aware and context-aware manner since algorithm executes in an environment in which the context often changes and computing resources are often severely limited. We propose to analyze the execution behavior of the data mining algorithm by mining its past executions. In order to extract the behavior model from algorithm's executions, we make use of two different data mining methods which are Bayesian network and decision tree classifier.
Bayesian network is constructed in order to represent the probabilistic relationships among device's resource usage, context, algorithm parameter settings and the performance of data mining. Based on this knowledge, future data mining algorithm configurations appropriate for situations are inferred.
Other data mining method that has been used to model the behavior of data mining algorithm, is the decision tree classifier. The effects of resource and context states as well as parameter settings on the data mining quality are discovered through decision tree classifier. In this approach, a taxonomy is defined on data mining quality so that tradeoff between prediction accuracy and classification specificity of each behavior model that classifies by a different abstraction of quality, is scored for model selection.
We formally define the behavior model constituents, instantiate the approach for association rules and validate the feasibility of the two of the approaches by the experimentation.