Music composition under traditional culture not only involves complex artistic expression, but also needs to be combined with modern computational methods to improve the efficiency and quality of composition. This study proposes a music composition model based on the Improved Multitrack Sequence Generation Adversarial Network (RFGAN), which aims to improve the quality and coherence of the generated music. The model optimizes the music generation process by introducing a loop-structured generator and timing model, combined with a discriminative feedback mechanism inside and outside the tracks. Comparative experiments were conducted to evaluate the model’s note prediction using Top1, Top2 and Top3 accuracies, and the results showed that RFGAN achieved the highest 88.79% Top1 accuracy in note prediction. To further validate the effectiveness of the model, the study also used a twelve-mean rhythm comparison, and the results showed that the generated note distributions were similar to the real music data, indicating that the model was able to capture the regularity of the music. In addition, the music generated by the model also outperforms the traditional GAN and BiGRU models in terms of harmony, rhythm, and overall effect, verifying its advantages in music composition.
