In Music Information Retrieval, classification of music genres is a core task and has been gaining increasing interest by adopting automated classification methods. Among these approaches, ensemble learning techniques have emerged as a promising solution by demonstrating their ability to enhance classification performance across diverse domains. However, traditional ensemble learning techniques may not deliver the desired accuracy improvements when applied to music datasets characterized by highly correlated low-level features associated with music genres. This study presents an innovative ensemble learning technique to address this challenge. The effectiveness of this approach is evaluated alongside established ensemble learning techniques by utilizing three publicly available music datasets, with two containing high-level sentiment-related features and one comprising low-level features extracted from music signals. The empirical experiments indicate that the proposed ensemble learning technique constantly outperforms conventional techniques in terms of classification accuracy. Notably, the proposed technique demonstrates remarkable performance enhancements when processing low-level features, whereas the traditional techniques failed to do so. This research highlights the substantial potential of advanced ensemble learning techniques in music genre classification and provides valuable insights into the strengths and limitations of various ensemble learning techniques when confronted with complex heterogeneous music datasets.