2021-10-12这看的是昨天的论文，RecSys21 Debiased Explainable Pairwise Ranking from Implicit Feedback

看了2篇论文了；1篇bias综述；2篇自动去偏模型；3篇基于内隐反馈的Debiase可解释成对排序；这篇难；看看就放弃了；

2 BACKGROUND
In this section, we start by reviewing previous work on explainabil-
ity and counteracting exposure bias in recommendation. While it is
impossible to do justice to every past contribution with an exhaus-
tive review, we try to focus on the most representative or related
work. Then we review Bayesian Personalized Ranking (BPR).

2背景在本节中，我们首先回顾以前关于解释性和抵消建议中暴露偏差的工作。虽然不可能通过详尽的审查来公正地对待过去的每一项贡献，但我们努力将重点放在最具代表性或最相关的工作上。然后，我们回顾了贝叶斯个性化排名（BPR）。

8 CONCLUSION
We proposed a novel explainable pairwise ranking loss with a corre-
sponding MF-based model called Explainable Bayesian Personalized
Ranking. We theoretically quantified the additional exposure bias
resulting from the explainability, and proposed an IPS-based unbi-
ased estimator for the ideal loss. We tested our proposed approaches
on three recommendation tasks and presented an extensive discus-
sion about the advantages of the proposed explainability extension;
as well as the impact of the debiasing, for varying data sparsities
and varying neighborhood sizes. Finally, we studied the popularity-
debiasing properties of the proposed methods in terms of Novelty,

Popularity, and Diversity; and unveiled an inherent popularity debi-
asing stemming from the neighborhood interactions. Our findings
are informative and motivate further research because our pro-
posed EBPR model yielded the best performance overall with no
significant trade-off between explainability and accuracy. Moreover,
we showed how combining explainability and exposure debiaing
yields powerful popularity debiasing through the proposed UEBPR
loss. Finally, our results point towards EBPR and pUEBPR being
the top performers that offer the best tradeoff between accuracy,
explainability and debiasing capacity. However, despite their com-
petitive performance, our proposed approaches may suffer from
the vanishing gradient problem in extremely sparse settings.

8结论我们提出了一种新的可解释的两两排序损失，以及相应的基于MF的模型，称为可解释的贝叶斯个性化排序。我们从理论上量化了可解释性导致的额外暴露偏差，并提出了一种基于IPS的理想损失无偏估计。我们在三个推荐任务上测试了我们提出的方法，并对提出的可解释性扩展的优势进行了广泛的讨论；对于不同的数据稀疏性和不同的邻域大小，以及debiasing的影响。最后，我们从新颖性的角度研究了建议方法的大众化偏差特性，大众化和多样性；并揭示了源自邻里互动的内在流行性。我们的研究结果提供了信息，并推动了进一步的研究，因为我们提出的EBPR模型在可解释性和准确性之间没有显著的权衡，总体上产生了最佳的性能。此外，我们还通过提议的UEBPR损失展示了可解释性和暴露性的结合如何产生强大的流行性借记。最后，我们的结果表明，EBPR和pUEBPR是在准确性、可解释性和借记能力之间提供最佳折衷的最佳执行者。然而，尽管我们提出的方法具有竞争性的性能，但在非常稀疏的环境中，我们提出的方法可能会遇到消失梯度问题。

总：这篇论文没看懂，主要是BPR不会，所以提出的新的模型EBPR也看不懂，但是我能知道新的EBPR好，能解决很多问题，但是也有一点不好，在非常稀疏的环境中，我们提出的方法可能会遇到消失梯度问题。

看不懂