- 推荐模型优化方案与实现
由于自然推荐场景的隐式反馈形式(1/0), 搜索场景的LTRLoss形式在自然推荐场景精排模型上有一定的限制性, 但是可以使用在粗排模型蒸馏精排模型的优化中。
- BPR Loss
- DValue-Weighted Distill Ranking Loss
- Bucket PairWise Ranking Loss
-
ListNet/ListMLE
-
LambdaLoss:
-
ApproxMetricLoss(NDCG/MRR)
- Multi-task learning using uncertainty to weigh losses for scene geometry and semantics.
- GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks
- End-to-End Multi-Task Learning with Attention
- 阿里-Rocket Launching:A Universal and Efficient Framework for Training Well-Performing Light Net
- 华为-IntTower: the Next Generation of Two-Tower Model for Pre-Ranking System
- 腾讯-Mixture of Virtual-Kernel Experts for Multi-Objective User Profile Modeling
- Google-Empowering Long-tail Item Recommendation through Cross Decoupling Network (CDN)
- 快手-POSO: Personalized Cold Start Modules for Large-scale Recommender Systems.