FC-MST: Feature correlation maximum spanning tree for multimedia concept classification

Feature selection is an actively researched topic in varies domains, mainly owing to its ability in greatly reducing feature space and associated computational time. Given the explosive growth of high-dimensional multimedia data, a well-designed feature selection method can be leveraged in classifying multimedia contents into high-level semantic concepts. In this paper we present a multi-phase feature selection method using maximum spanning tree built from feature correlation among multiple modalities (FC-MST). The method aims to first thoroughly explore not only the correlation between features within and across modalities, but also the association of features towards semantic concepts. Secondly, with the correlations, we identify important features and exclude redundant or irrelevant ones. The proposed method is tested on a well-known benchmark multimedia data set called NUS-WIDE and the experimental results show that it outperforms four well-known feature selection methods in all three important measurement metrics.