An Univariate Feature Elimination Strategy for Clustering Based on Metafeatures
Saptarsi Goswami, Sanjay Chakraborty, Himadri Nath Saha · International Journal of Intelligent Systems and Applications · 2017
Feature selection plays a very important role in all pattern recognition tasks.It has several benefits in terms of reduced data collection effort, better interpretability of the models and reduced model building and execution time.A lot of problems in feature selection have been shown to be NP -Hard.There has been significant research in feature selection in last three decades.However, the problem of feature selection for clustering is still quite an open area.The main reason is unavailability of target variable as compared to supervised tasks.In this paper, five properties or metafeatures like entropy, skewness, kurtosis, coefficient of variation and average correlation of the features have been studied and analysed.An extensive study has been conducted over 21 publicly available datasets, to evaluate viability of feature elimination strategy based on the values of the metafeatures for feature selection in clustering.A strategy to select the most appropriate metafeatures for a particular dataset has also been outlined.The results indicate that the performance decrease is not statistically significant.