原文发表于:2020-12-14

解决方案

position作为特征

训练是加入位置信息

预估时设置为默认值

缺点:会造成非默认值的预估偏差

Shallow Tower

shallow tower建模偏差特征

训练时随机drop掉

预测时直接丢掉shallow tower

PAL

https://www.researchgate.net/publication/335771749_PAL_a_position-bias_aware_learning_framework_for_CTR_prediction_in_live_recommender_systems

DPIN

https://arxiv.org/pdf/2106.05482.pdf

参考

  • PAL: a position-bias aware learning framework for CTR prediction in live recommender systems
  • https://zhuanlan.zhihu.com/p/145448160?from_voters_page=true
  • 美团DPIN https://blog.csdn.net/m0_47256162/article/details/125527968