Correlation-Based Feature Selection and Regression

Music video is a well-known medium in music entertainment which contains rich affective information and has been widely accepted as emotion expressions. Affective analysis plays an important role in the content-based indexing and retrieval of music video. This paper proposes a general scheme for music video affective estimation using correlation-based feature selection followed by regression. Arousal score and valence score with four grade scales are used to measure music video affective content in 2D arousal/valence space. The main contributions are in the following aspects: (1) correlation-based feature selection is performed after feature extraction to select representative arousal and valence features; (2) different regression methods including multiple linear regression and support vector regression with different kernels are compared to find the fittest estimation model. Significant reductions in terms of both mean absolute error and variation of absolute error compared with the state-of-the-art methods clearly demonstrate the effectiveness of our proposed method.