Reproduce “Multi-Faceted Hierarchical Multi-Task Learning for a Large Number of Tasks with Multi-dimensional Relations” on OpenResearch

About this paper

Junning Liu, Zijie Xia, Yu Lei, Xinjian Li, Xu Wang

Multi-Faceted Hierarchical Multi-Task Learning (MFH) organizes large task collections around explicit dimensions such as user behavior, user activity, and item status, rather than connecting all tasks through a flat multi-task architecture. Evaluations on roughly 10 billion WeSee short-video samples, online recommendation traffic, and Ali-CCP show improved prediction for data-sparse groups, including up to 9.1% higher app time per new user and 1.85% higher next-day retention.

Key insights

• Increasing the number of tasks alone is insufficient; the way task relations are organized has a larger effect than simply adding task-specific towers to a flat model.
• Intersecting hierarchies provide specialized representations for sparse regions while retaining transfer from related, better-sampled tasks.
• The architecture is especially useful when task dimensions correspond to meaningful business or behavioral structure, such as user activity, content maturity, and interaction type.

Summary by alphaXiv.