Effective music skip prediction based on late fusion architecture for user-interaction noise

초록

Music skip prediction aims to predict whether user skips occur in upcoming songs during a playlist streaming session. The track features representing musical characteristics, such as danceability, loudness, and key, and session features containing user interaction logs can be fused to improve prediction performance. However, existing music skip prediction methods fuse two information sources without considering the noise of session features, degrading prediction performance. This study proposes a new neural network that extracts essential information from each source and then fuses information for skip prediction. Based on the proposed architecture, a better quality of fused information can be exploited because the noise can be removed before fusion. The proposed method was compared with conventional neural networks, and the experiments revealed that the proposed method significantly outperforms the compared methods. © 2023 Elsevier Ltd

키워드