Multi-Faceted Hierarchical Multi-Task Learning for a Large Number of Tasks with Multi-dimensional Relations

• DOI: 10.48550/arxiv.2110.13365
• Publication Date: 2021-01-01
• Authors: Junning Liu; Zijie Xia; Yu Lei; Xinjian Li; Xu Wang
• Keywords: Overfitting

There has been many studies on improving the efficiency of shared learning in Multi-Task Learning(MTL). Previous work focused "micro" sharing perspective for a small number tasks, while Recommender Systems(RS) and other AI applications, there are often demands to model large tasks with multi-dimensional task relations. For example, when using MTL various user behaviors RS, if we differentiate new users items from old ones, will be cartesian product style increase This "macro" network design proposes Multi-Faceted Hierarchical model(MFH). MFH exploits multi-dimension relations nested hierarchical tree structure which maximizes learning. We evaluate SOTA models industry video platform 10 billion samples results show that outperforms significantly both offline online evaluations across all groups, especially remarkable an 9.1\% app time per 1.85\% next-day retention rate. now deployed scale recommender system. is beneficial cold-start problems RS where suffer "local overfitting" phenomenon. However, idea actually generic widely applicable scenarios.