Optimizing parameters values of tree-based contrast subspace Miner using genetic algorithm

• Main Authors: Florence Sia, Rayner Alfred
• Format: Conference or Workshop Item
• Language: English; English
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: https://eprints.ums.edu.my/id/eprint/25530/7/Optimizing%20parameters%20values%20of%20tree-based%20contrast-Abstract.pdf; https://eprints.ums.edu.my/id/eprint/25530/1/Optimizing%20parameters%20values%20of%20tree-based%20contrast%20subspace%20Miner%20using%20genetic%20algorithm.pdf; https://eprints.ums.edu.my/id/eprint/25530/

Mining contrast subspace finds contrast subspaces or subspaces where a query object is most similar to a target class but different from other class in a two-class multidimensional data set. Tree-based contrast subspace miner (TB-CSMiner) which employs tree-based likelihood contrast scoring function has been recently introduced to mine contrast subspaces of a query object by constructing tree from a subspace that is data objects in a subspace space are divided into two nodes recursively with respect to the query object until the node contains only objects of same class or a minimum number of objects. A query object should fall in the node that has higher number of objects belong to the target class against the other class in a contrast subspace. The effectiveness of TB-CSMiner in finding contrast subspace of a query object relies on the values of several parameters involved which include the minimum number of objects in a node, the denominator of tree-based likelihood contrast scoring function, the number of relevant features for tree construction, and the number of random subspaces for contrast subspace search. It is difficult to identify the values of these parameters in a straightforward way based on the conventional analysis. As a consequence, this paper proposes a genetic algorithm based method for identifying the parameters values of TB-CSMiner in which sets of parameters values are treated as individuals and evolved to return the best set of parameters values. The experiment results show that the TB-CSMiner with parameters values identified through the genetic algorithm outperformed those identified through the conventional analysis in most of the cases.