推荐系统的模型训练经历了一个User→Data→Model→User的循环,在整个循环的过程中,会引入7种偏差(biases)。Debias技术有篇比较出名的Survey文章《Bias and Debias in Recommender System: A Survey and Future》对这7种偏差进行了解释:
这个方案主要利用validate数据集的二次训练解决over-confidence以及辅助优化目标导致的训练有偏问题。首先是Over Confidence的解决方案。有篇经典文章《On Calibration of Modern Neural Networks》认为在于复杂模型训练后模型准确率和置信度不匹配所致,因此在validate集合上基于Negtive Log likehood训练Softmax的Temperature参数
1. Bias and Debias in Recommender System: A Survey and Future
2. On Calibration of Modern Neural Networks
3. MBCT: Tree-Based Feature-Aware Binning for Individual Uncertainty Calibration
4. Distilled CTR: Accurate and scalable CTR prediction model through model distillation
5. Learning to Rank: From Pairwise Approach to Listwise Approach
6. Recommending What Video to Watch Next: A Multitask Ranking System
7. Generalized Loss Functions for Generative Adversarial Networks 8. PAL: a position-bias aware learning framework for CTR prediction in live recommender systems
9. Recommendations as treatments: Debiasing learning and evaluation.
10. Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
11. Adversarial Gradient Driven Exploration for Deep Click-Through Rate Prediction