Effective Feature Space Reduction with Imbalanced Data for Semantic Concept Detection

Semantic understanding of multimedia content has become a very popular research topic in recent years. Semantic concept detection algorithms face many challenges such as the semantic gap and imbalance data, among others. In this paper, we propose a novel algorithm using multiple correspondence analysis (MCA) to discover the correlation between features and classes to reduce the feature space and to bridge the semantic gap. Moreover, the proposed algorithm is able to explore the correlation between items (i.e., feature-value pairs generated for each of the features) and classes which expands its ability to handle imbalance data sets. To evaluate the proposed algorithm, we compare its performance on semantic concept detection with several existing feature selection methods under various well-known classifiers using some of the concepts and benchmark data available from the TRECVID project. The results demonstrate that our proposed algorithm achieves promising performance, and it performs significantly better than those feature selection methods in the comparison for the imbalanced data sets.