Uroboros0313 / RecCheatSheet

RecModuleCheatSheet

RecCheatSheet

• 推荐模型优化方案与实现

RankingLosses

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。

PairWise

• BPR Loss
• DValue-Weighted Distill Ranking Loss
• Bucket PairWise Ranking Loss
  • 美团-美团搜索粗排优化的探索与实践
  • 阿里-全链路联动-面向最终目标的全链路一致性建模

ListWise

• ListNet/ListMLE

  • ICML-2007-Learning to Rank: From Pairwise Approach to Listwise Approach

• LambdaLoss:

  • CIKM-2018-The LambdaLoss Framework for Ranking Metric Optimization
  • 大众点评搜索基于知识图谱的深度学习排序实践

• ApproxMetricLoss(NDCG/MRR)

  • 微软-A General Approximation Framework for Direct Optimization of Information Retrieval Measures

多Loss联合蒸馏

MTL多Loss学习

• 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

精排-粗排联合训练/共享底层/Embedding蒸馏

• 阿里-Rocket Launching：A Universal and Efficient Framework for Training Well-Performing Light Net

双塔模型结构优化

特征交叉

• 华为-IntTower: the Next Generation of Two-Tower Model for Pre-Ranking System

专家混合网络

• 腾讯-Mixture of Virtual-Kernel Experts for Multi-Objective User Profile Modeling
• Google-Empowering Long-tail Item Recommendation through Cross Decoupling Network (CDN)

千人千模

• 快手-POSO: Personalized Cold Start Modules for Large-scale Recommender Systems.