A Novel Feature Selection Algorithm for Strongly Correlated Attributes Using Two-Dimensional Discriminant Rules

Considerable attention has been devoted to the development of feature selection algorithms for various applications in the last decade. Most of them concentrate to the single attributes. In contrast, limited research work has been devoted to determine correlated and pairwise attributes or features due to the difficulty of the problem. We present a novel feature selection algorithm for strongly correlated attributes using the two-dimensional discriminant rules. We discover a subset of pairwise attributes whose target class is influenced not only by a single cause in numeric-based datasets, both categorical and continuous target classes. Our algorithm uses x-monotone optimization for determination of optimal region within datasets. For selection strategy, the region evaluation has been applied using skewness and kurtosis metric. The results are then compared to other numeric based attribute selection algorithms. The result shows a unique capability to reveal the importance of pairwise strongly correlated attributes that conventional methods missed to explore. Keyword strongly correlated attributes, two-dimensional discriminant rules approach