Abstract:
This study develops a music recommendation system based on user preferences, emphasizing
genre classification and playlist generation. The classification task is addressed using a self-created, bal anced dataset comprising 500 audio tracks, each 30 seconds in duration, equally distributed across five distinct Tamil music genres: classical, devotional, melody, romantic, and rock. Mel-Frequency Cepstral Coefficients (MFCCs) were extracted from the audio data, followed by data preprocessing techniques in cluding Label Encoding and Standard Scaling (Z-score normalization). To improve computational efficiency, dimensionality was reduced using Kernel Principal Component Analysis (KPCA). The processed dataset was then split in an 80:20 ratio for training and testing. Gradient Boosting models, specifically XGBoost and LightGBM, were employed, with andomizedSearchCV used for hyperparameter tuning. Initial train ing on the high-dimensional feature set yielded low accuracies (42% for XGBoost and 43% for LightGBM) with extended training times. Incorporating KPCA significantly improved both performance and efficiency: KPCA-enhanced XGBoost achieved 82% accuracy in 16 minutes, while LightGBM reached 86% in 18 min utes. In comparison, CNN, RNN, CRNN, Random Forest, and CNN+XGBoost models achieved lower
accuracies of 45.5%, 32%, 23%, 47%, and 52%, respectively. These results demonstrate that Gradient Boost ing algorithms, particularly LightGBM, outperform traditional machine learning and deep learning models, making them highly suitable for Tamil music genre classification and recommendation systems.
