Sparsity-Aware Personalized Pattern Extractor Network for Music Multi-task Learning

Shijia Wang, Yi Zheng, Qiang Xiao, Yilong Zhao, Qimeng Yang, Chuanjiang Luo

Abstract: Music recommendation can be modeled as ranking problems based on multi-task estimation. These tasks can be categorized into two main parts: the click-through rate estimation task (CTR), which involves user behaviors such as playing or clicking on music resources, and the conversion rate estimation task (CVR), which encompasses collecting and sharing music. The effectiveness of the typical ranking criterion, the product of CTR and CVR, is often affected by the relative sample scale and sparsity between the two estimation tasks. This disparity in sparsity can lead to suboptimal ranking results and negatively impact the overall performance of the recommender system. In online music recommendation scenario, we observed some practical issues such as overexposure of eye-catching playlists and poor transferability of the model between different tasks. In this paper, we first utilize the joint distribution of CTR and CVR to quantify the relative disparity in sparsity between the two tasks. To make better use of the aforementioned joint distribution and guide the fusion of CTR and CVR, we propose a Sparsity-Aware Personalized Pattern Extractor Network (PPEN) to provide fine-grained information at the instance level, which generates adjustment parameters according to specific patterns learned in an end-to-end fashion. Extensive offline experiments are conducted to evaluate the effectiveness of our proposed method on both industrial and public datasets. Additionally, our model has been successfully deployed on Netease Cloud Music, one of China’s largest music streaming platforms, and has shown significant improvements in A/B testing.