Reproduce “Deep Task-specific Bottom Representation Network for Multi-Task Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qi Liu, Zhilong Zhou, Gangwei Jiang, Tiezheng Ge, Defu Lian
Deep Task-specific Bottom Representation Network (DTRN) addresses negative transfer in multi-task recommendation by learning separate task-conditioned representations before the shared prediction layers. Across a 376-million-sample e-commerce dataset and Tenrec, it achieved the best reported AUC and LogLoss among the evaluated baselines, while an online test improved watch, click, entering, and revenue-per-mille rates by 0.63%, 1.00%, 2.23%, and 1.00%, respectively.
Key insights
- Task conditioning is useful before, not only within, the multi-task prediction network; separating bottom representations can reduce interference between objectives with different feature preferences.
- Conditional layer normalization was more effective overall than injecting task information into Transformer attention or feed-forward layers, while remaining parameter-efficient compared with fully independent sequence models.
- Ablations show that task-specific interest extraction and task-specific feature refinement provide complementary gains, and task relationships can be asymmetric—for example, watch behavior supports later actions more than later actions support watching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.