作者:袁冬至
微信:DataGap
公众号:救命的药
研究方向:推荐系统
欢迎交流,学习!
RecSys21 Debiased Explainable Pairwise Ranking from Implicit Feedback
又是一篇新的文章;看;
发到公众号是必须的,个人影响力;发到网上必须的大家可以搜索到,公众号以前的文章没人看;
Debiased Explainable Pairwise Ranking from Implicit Feedback
基于内隐反馈的Debiase可解释成对排序
ABSTRACT
Recent work in recommender systems has emphasized the impor-
tance of fairness, with a particular interest in bias and transparency,
in addition to predictive accuracy. In this paper, we focus on the
state of the art pairwise ranking model, Bayesian Personalized
Ranking (BPR), which has previously been found to outperform
pointwise models in predictive accuracy, while also being able to
handle implicit feedback. Specifically, we address two limitations
of BPR: (1) BPR is a black box model that does not explain its out-
puts, thus limiting the user’s trust in the recommendations, and
the analyst’s ability to scrutinize a model’s outputs; and (2) BPR
is vulnerable to exposure bias due to the data being Missing Not
At Random (MNAR). This exposure bias usually translates into an
unfairness against the least popular items because they risk be-
ing under-exposed by the recommender system. In this work, we
first propose a novel explainable loss function and a corresponding
Matrix Factorization-based model called Explainable Bayesian Per-
sonalized Ranking (EBPR) that generates recommendations along
with item-based explanations. Then, we theoretically quantify ad-
ditional exposure bias resulting from the explainability, and use it
as a basis to propose an unbiased estimator for the ideal EBPR loss.
The result is a ranking model that aptly captures both debiased
and explainable user preferences. Finally, we perform an empirical
study on three real-world datasets that demonstrate the advantages
of our proposed models.
最近在推荐系统中的工作强调了公平的重要性,除了预测准确性外,还特别关注偏见和透明度。在本文中,我们关注最新的成对排序模型,即贝叶斯个性化排序(BPR),该模型在预测精度方面优于点式模型,同时也能够处理隐式反馈。具体来说,我们解决了BPR的两个局限性:(1)BPR是一个黑箱模型,它不解释其输出,因此限制了用户对建议的信任,以及分析师审查模型输出的能力;(2)由于数据丢失而非随机(MNAR),BPR容易受到暴露偏差的影响。这种曝光偏差通常转化为对最不受欢迎的项目的不公平,因为它们有被推荐系统曝光不足的风险。在这项工作中,我们首先提出了一种新的可解释损失函数和相应的基于矩阵分解的模型,称为可解释贝叶斯个性化排名(EBPR),该模型生成建议以及基于项目的解释。然后,我们从理论上量化了可解释性导致的额外暴露偏差,并以此为基础提出了理想EBPR损失的无偏估计量。结果是一个排名模型,它恰当地捕获了借记和可解释的用户偏好。最后,我们对三个真实数据集进行了实证研究,证明了我们提出的模型的优势。
总:逐点模型不好,贝叶斯个性化排序(BPR)好,但是有2个地方不好,我们能改进这个模型提出了可解释贝叶斯个性化排名(EBPR),实验证明它好;
1 INTRODUCTION
Bayesian Personalized Ranking (BPR) is a state of the art pairwise
ranking approach [36] that has recently received significant praise
in the recommender systems community because of its capacity
to rank implicit feedback data with high accuracy compared to
pointwise models [18]. Aiming to rank relevant items higher than
irrelevant items, pairwise ranking recommender systems often
assume that all non-interacted items as irrelevant. Hence, these
systems rely on the assumption that implicit feedback data is Miss-
ing Completely At Random (MCAR), meaning that the items are
equally likely to be observed by the users [40], consequently any
missing interaction is missing because the user chose not to in-
teract with it. However, given the abundance of items on most
e-commerce, entertainment, and other online platforms, it is safe to
assume the impossibility of any user being exposed toallthe items.
Thus, missing interactions should be considered Missing Not At
Random (MNAR). This means that the user may have been exposed
to part of the items, but chose not to interact with them, which can
be a sign of irrelevance; and was not exposed to the rest of the items.
This MNAR property is translated into an exposure bias. This type
of bias is usually characterized by a bias against less popular items
that have a lower propensity of being observed [6].
贝叶斯个性化排名(BPR)是一种最先进的成对排名方法[36],最近在推荐系统界受到了极大的赞扬,因为与逐点模型相比,它能够以高精度对隐式反馈数据进行排名[18]。两两排序推荐系统的目标是将相关项目的排名高于不相关项目,通常假设所有未交互的项目都是不相关的。因此,这些系统依赖于隐式反馈数据完全随机缺失的假设(MCAR),这意味着这些项目同样可能被用户观察到[40],因此任何缺失的交互都会缺失,因为用户选择不与之交互。然而,考虑到大多数电子商务、娱乐和其他在线平台上有大量的物品,可以放心地假设任何用户都不可能接触到所有这些物品。因此,缺失的交互作用应视为缺失而非随机(MNAR)。这意味着用户可能接触到部分项目,但选择不与它们交互,这可能是无关的迹象;并且没有接触到其他物品。此MNAR特性转化为曝光偏差。这种类型的偏见通常表现为对不太受欢迎的项目的偏见,这些项目被观察的倾向性较低[6]。
Moreover, most accurate recommender systems tend to be black
boxes that do not justify why or how an item was recommended to
a user. This might engender unfairness issues if, for example, par-
ticularly inappropriate or offensive content gets recommended to a
user. This kind of unfairness can be better diagnosed and mitigated
with an explanation. In fact, it could be important for the user to
know why or how the inappropriate item was recommended. For
example, an Italian user might think that the movie recommen-
dation “The Godfather" is offensive because of the way it depicts,
in an unfair stereotypical way, a certain Italian community in the
US. However, the explanation “Because you liked the movie “Scar-
face"" can be important in this case, because it clarifies that the
movie recommendation was not tied to a community, but rather
resulted from the user also liking another similar “mafia" sub-genre
movie. Furthermore, explanations have been shown to help users
make more accurate decisions, which translates into an increased
user satisfaction [2,4]. Bayesian Personalized Ranking [36] treats
comparisons between any positive and negative items the same, re-
gardless of which ones can or cannot be explained. Thus, while BPR
aptly captures and models ranking based preference, it does not yet
capture anexplainablepreference. It is thisexplainable preference, in
addition to an unbiased preference ranking, that we seek to achieve
in this work. We thus propose models that address explainability
andexposure bias in pairwise ranking from implicit feedback and
achieve the following contributions:
•Proposing an explainable loss function based on the state of
the art Bayesian Personalized Ranking (BPR) loss [36] along
with a corresponding Matrix Factorization (MF)-based model
called Explainable Bayesian Personalized Ranking (EBPR).
To the extent of our knowledge, no work has introduced
neighborhood-based explainability to pairwise ranking.
•Conducting a theoretical study of the additional exposure
bias coming from the item-based explanations.
•Proposing an unbiased estimator for the ideal EBPR loss,
called UEBPR, based on the Inverse Propensity Scoring (IPS)
estimator [37]. To our knowledge, no prior work has tried to
address the additional exposure bias that could result from
neighborhood-based explainability.
•Performing an empirical study on three real-world datasets
to compare the effectiveness of the proposed models, in terms
of ranking, explainability, and both exposure and popularity
debiasing.
•Investigating the properties of the proposed neighborhood
based explainable models, revealing and explaining a desir-
able inherent popularity debiasing that is built into these
models. This opens the path to a new family of future debias-
ing strategies, where the debiasing is rooted in an explainable
neighborhood-based rationale.
In addition, we make our implementations of all the models
presented in this paper available for reproducibility1.
此外,最准确的推荐系统往往是黑匣子,无法证明向用户推荐商品的原因或方式。例如,如果向用户推荐特别不合适或冒犯性的内容,这可能会产生不公平问题。这种不公平可以通过解释得到更好的诊断和缓解。事实上,对于用户来说,了解为什么或如何推荐不合适的项目可能很重要。例如,一位意大利用户可能认为电影推荐“教父”是冒犯性的,因为它以一种不公平的陈规定型方式描述了美国的某个意大利社区。然而,解释“因为你喜欢电影“疤脸”“在这种情况下可能很重要,因为它澄清了电影推荐与社区无关,而是因为用户也喜欢另一部类似的“黑手党”子类型电影。此外,还提供了一些解释,帮助用户做出更准确的决定,从而提高了用户满意度[2,4].贝叶斯个性化排名[36]将任何积极和消极项目之间的比较视为相同的,无论哪些项目可以解释或无法解释。因此,尽管BPR恰当地捕获并建模了基于排名的偏好,但它尚未捕获可解释的参考。在除了一个无偏见的偏好排名,我们在这项工作中寻求实现。因此,我们提出了模型,从隐式反馈解决两两排序中的可解释性和暴露偏见,并取得以下贡献
基于最先进的贝叶斯个性化排名(BPR)损失[36]以及相应的基于矩阵分解(MF)的模型,提出了一个可解释损失函数,称为可解释贝叶斯个性化排名(EBPR)。就我们所知,没有任何工作将基于邻域的解释引入成对排序对基于项目的解释产生的额外暴露偏差进行理论研究基于逆倾向评分(IPS)估计器[37],提出理想EBPR损失的无偏估计器,称为UEBPR。据我们所知,之前没有任何工作试图解决基于邻里关系的可解释性可能导致的额外暴露偏差对三个真实世界的数据集进行实证研究,比较所提出模型在排名、可解释性、曝光率和受欢迎度方面的有效性调查所提出的基于邻域的可解释模型的性质,揭示和解释这些模型所固有的可取的普遍性。这为未来的借记策略开辟了一条新的家族之路,其中的借记植根于可解释的基于邻域的原理。此外,我们使本文中提出的所有模型的实现可用于再现性1。