A Novel Feature Selection Measure Partnership-Gain


  • Mostafa A. Salama British University in Egypt - BUE
  • Ghada Hassan British University in Egypt - BUE




Feature Selection, Interdependency between features, Classification


Multivariate feature selection techniques search for the optimal features subset to reduce the dimensionality and hence the complexity of a classification task. Statistical feature selection techniques measure the mutual correlation between features well as the correlation of each feature to the tar- get feature. However, adding a feature to a feature subset could deteriorate the classification accuracy even though this feature positively correlates to the target class. Although most of existing feature ranking/selection techniques consider the interdependency between features, the nature of interaction be- tween features in relationship to the classification problem is still not well investigated. This study proposes a technique for forward feature selection that calculates the novel measure Partnership-Gain to select a subset of features whose partnership constructively correlates to the target feature classification. Comparative analysis to other well-known techniques shows that the proposed technique has either an enhanced or a comparable classification accuracy on the datasets studied. We present a visualization of the degree and direction of the proposed measure of features’ partnerships for a better understanding of the measure’s nature.




How to Cite

Salama, M. A., & Hassan, G. (2019). A Novel Feature Selection Measure Partnership-Gain. International Journal of Online and Biomedical Engineering (iJOE), 15(04), pp. 4–19. https://doi.org/10.3991/ijoe.v15i04.9831