Comparative Analysis of Music Mood Classification Methods

Music mood classification is a significant research area in the field of music information retrieval (MIR). Understanding and categorizing the emotional content of music is crucial for various applications. Numerous studies on music classification have been performed. Although most of these studies reports high accuracy levels, however, the overfitting problem may exist. This paper performs an exploratory analysis to overcome overfitting problems and investigates approaches towards improving the performance of music mood classification system (MMCS). Initially, the 4Q emotion dataset was used with input acoustic features in 1D and 2D formats. Different CNN architectures were designed with parameter optimizations for the 1D and 2D data formats. Results from the initial experiment clearly indicated the overfitting problem. Following this, a comparative evaluation was then performed with the 4Q Turkish music dataset. Further exploratory analysis was conducted. Data augmentation techniques were applied to overcome insufficient data. The results of this analysis are presented.