Paper Type
Complete
Paper Number
PACIS2025-1767
Description
User engagement matters to the attention economy, such as the short-video platform, but predicting multiple engagement outcomes simultaneously is challenging due to complex interdependencies and data sparsity, often causing the negative transfer or seesaw effect in traditional multi-task learning. We propose the Mutual Information–enhanced Multi-task Learning (MIML) framework, integrating a mixture-of-experts design and GCNs to learn task-specific and shared feature with social propagation of preferences. By incorporating a conditional mutual information term into its loss function, which is approximated by a mutual information proxy, MIML quantifies and leverages task correlations for information sharing. Theoretical analysis validates the effectiveness of mutual information. Experiments on simulation data and two public short-video datasets demonstrate the efficiency of mutual information and our proposed method, which means that MIML surpasses state-of-the-art methods and achieve higher AUCs with mitigating the seesaw effect, thereby offering practical insights for optimizing content delivery and boosting user engagement.
Recommended Citation
Yao, Xinpei; Wang, Cong; and Guo, Xunhua, "Decoding User Engagement: A Mutual Information-Enhanced Multi-Task Learning Framework for Predicting Engagement on Social Media" (2025). PACIS 2025 Proceedings. 20.
https://aisel.aisnet.org/pacis2025/aiandml/aiandml/20
Decoding User Engagement: A Mutual Information-Enhanced Multi-Task Learning Framework for Predicting Engagement on Social Media
User engagement matters to the attention economy, such as the short-video platform, but predicting multiple engagement outcomes simultaneously is challenging due to complex interdependencies and data sparsity, often causing the negative transfer or seesaw effect in traditional multi-task learning. We propose the Mutual Information–enhanced Multi-task Learning (MIML) framework, integrating a mixture-of-experts design and GCNs to learn task-specific and shared feature with social propagation of preferences. By incorporating a conditional mutual information term into its loss function, which is approximated by a mutual information proxy, MIML quantifies and leverages task correlations for information sharing. Theoretical analysis validates the effectiveness of mutual information. Experiments on simulation data and two public short-video datasets demonstrate the efficiency of mutual information and our proposed method, which means that MIML surpasses state-of-the-art methods and achieve higher AUCs with mitigating the seesaw effect, thereby offering practical insights for optimizing content delivery and boosting user engagement.
Comments
AI ML