10/11论文关键词，自动去偏框架论文翻译，发文章思考

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！

AutoDebias: Learning to Debias for Recommendation

还是看的这篇文章；

ABSTRACT 摘要

CCS CONCEPTS CCS的概念

KEYWORDS 关键词

1 INTRODUCTION 引言

2 PRELIMINARY 准备

2.1 Task Formulation2.1任务制定

2.2 Biases in Recommendation 2.2推荐偏差

3 A GENERAL DEBIASING FRAMEWORK 一个通用的去偏框架

3.1 Debiasing Empirical Risk 3.1去偏经验风险

3.2 Link to Related Work 3.2相关工作链接

4 AUTODEBIAS ALGORITHM 4 AUTODEBIAS算法

5 THEORETICAL ANALYSIS 5理论分析

6 EXPERIMENTS 6实验

6.1 Experimental Setup6.1实验装置

6.2 Performance Comparison on Explicit
Feedback (RQ1)6.2显式反馈的绩效比较(RQ1)

6.3 Ablation Study (RQ2)6.3消融研究(RQ2)

6.4 Exploratory Analysis (RQ3)6.4探索性分析(RQ3)

6.5 Performance Comparison on Implicit and
Feedback on Lists (RQ4)6.5表隐式与表反馈的表现比较(RQ4)

7 RELATED WORK7相关工作

8 CONCLUSION AND FUTURE WORK结论与未来工作

ACKNOWLEDGMENTS感谢

REFERENCES参考文献

ABSTRACT

Recommender systems rely on user behavior data like ratings and
clicks to build personalization model. However, the collected data
is observational rather than experimental, causing various biases in
the data which significantly affect the learned model. Most existing
work for recommendation debiasing, such as the inverse propensity
scoring and imputation approaches, focuses on one or two specific
biases, lacking the universal capacity that can account for mixed or
even unknown biases in the data.

推荐系统依赖于用户行为数据，如评分和点击来建立个性化模型。然而，收集的数据是观察性的，而不是实验性的，导致数据中的各种偏差，显著影响学习模型。大多数现有的推荐去偏方法，如反倾向评分和归因方法，侧重于一两个特定的偏差，缺乏能够解释数据中混合甚至未知偏差的普遍能力。

Towards this research gap, we first analyze the origin of bi-
ases from the perspective ofrisk discrepancythat represents the
difference between the expectation empirical risk and the true
risk. Remarkably, we derive a general learning framework that
well summarizes most existing debiasing strategies by specifying
some parameters of the general framework. This provides a valu-
able opportunity to develop a universal solution for debiasing,
e.g., by learning the debiasing parameters from data. However,
the training data lacks important signal of how the data is biased
and what the unbiased data looks like. To move this idea forward,
we proposeAotoDebiasthat leverages another (small) set of uni-
form data to optimize the debiasing parameters by solving the
bi-level optimization problem with meta-learning. Through theo-
retical analyses, we derive the generalization bound for AutoDebias
and prove its ability to acquire the appropriate debiasing strategy.
Extensive experiments on two real datasets and a simulated dataset
demonstrated effectiveness of AutoDebias. The code is available at
https://github.com/DongHande/AutoDebias.

针对这一研究差距，我们首先从风险差异的角度分析了偏差的起源，风险差异代表了预期经验风险和真实风险之间的差异。值得注意的是，我们导出了一个通用的学习框架，它通过指定通用框架的一些参数很好地总结了大多数现有的去偏策略。这提供了一个宝贵的机会，以发展一个普遍的解去偏，例如，通过学习去偏参数的数据。然而，训练数据缺乏重要的信号，表明数据是如何有偏差的，以及无偏差的数据是什么样的。为了推进这个想法，我们提出了aotodebias，利用另一个(小)统一数据集来优化去偏参数，通过用元学习解决双层优化问题。通过理论分析，我们推导了AutoDebias的泛化界，并证明了它能够获得合适的去偏策略。在两个真实数据集和一个模拟数据集上的大量实验证明了AutoDebias的有效性。代码可在https://github.com/DongHande/AutoDebias上获得。

风险差异；元学习；

这篇文章主要就是用元学习解决各种偏差，自动学习各个偏差弄成一个框架；

8 CONCLUSION AND FUTURE WORK
This paper develops a universal debiasing framework that not only
can well account for multiple biases and their combinations, but also
frees human efforts to identify biases and tune the configurations.
We first formulate various data biases as a case of risk discrepancy,
and then derive a general learning framework that subsumes most
debiasing strategies. We further propose a meta-learning-based
algorithm to adaptively learn the optimal debiasing configurations
from uniform data. Both theoretical and empirical analyses have
been conducted to validate the effectiveness of our proposal.
One interesting direction for future work is to explore more so-
phisticate meta model, which could capture more complex patterns
and potentially achieve better performance than linear meta model.
Also, note that in real-world biases are usually dynamic rather than
static. it will be valuable to explore how bias evolves with the time
goes by and develop a universal solutions for dynamic biases.

本文开发了一个通用的去偏框架，它不仅可以很好地解释多种偏差及其组合，而且还可以解放人类的努力来识别偏差和调整配置。我们首先将各种数据偏差作为风险差异的一个例子，然后得出一个包含大多数去偏策略的一般学习框架。我们进一步提出了一种基于元学习的自适应学习最优去偏配置的算法。通过理论分析和实证分析，验证了该方案的有效性。未来工作的一个有趣方向是探索更复杂的元模型，它可以捕获更复杂的模式，并可能获得比线性元模型更好的性能。另外，请注意，在现实世界中，偏见通常是动态的，而不是静态的。探索偏差如何随着时间的推移而演变，并为动态偏差制定一个通用的解决方案，将是很有价值的。

未来研究点：

1 更复杂的元模型，捕获更复杂的模式，并可能获得比线性元模型更好的性能。

2 动态偏差；探索偏差如何随着时间的推移而演变，并为动态偏差制定一个通用的解决方案，将是很有价值的。

7 RELATED WORK
Bias in recommendation.Besides the data bias that has been
detailed in section 2&3, two important biases in recommendation
results have been studied: (1) When a recommender model is trained
on a long-tailed data, popular items are recommended even more
frequently than their popularity would warrant [2], raising so-called
popularity bias. The long-tail phenomenon is common in RS data,
and ignoring the popularity bias will incur many issues,e.g., hurt-
ing user serendipity, making popular items even more popular. To
deal with this problem, [1,11,49] introduced regularizers to guide
the model to give more balanced results; [50,54] disentangled the
effect of user interest and item popularity with causal inference to
address popularity bias. (2) The system systematically and unfairly
discriminates against certain individuals or groups of individuals
in favor others, raising so-calledunfairness. Unfairness happens as
different user (or item) groups are usually unequally represented in
data. When training on such unbalanced data, the models are highly
likely to learn these over-represented groups, and potentially dis-
criminates against other under-represented groups [13,43]. There
are four types of strategies on addressing unfairness, including
re-balancing[41], regularizer[51], adversarial learning [4], causal
inference [26]. We encourage the readers refer to the survey [5] for
more details.

Meta learning in recommendation.Meta learning is an auto-
matic learning algorithms that aims at using metadata to improve
the performance of existing learning algorithms or to learn (induce)
the learning algorithm itself [14,20,46]. There is also some work in-
troducing meta learning in RS. For example, [10,36] leveraged meta
learning to learn a finer-grained regularization parameters; [53]
leveraged meta-learning to guide the re-training of a recommen-
dation model towards better performance; Meta-learning also has
been utilized to address the cold-start problem [30].

推荐中的相关工作偏见。除了在第2节和第3节中详细介绍的数据偏差外，还研究了推荐结果中的两个重要偏差:(1)当推荐模型在长尾数据上训练时，热门项目被推荐的频率甚至超过其受欢迎程度所保证的[2]，从而产生所谓的受欢迎偏差。长尾现象在RS数据中很常见，忽略受欢迎度偏差会导致许多问题，例如:，损害用户的意外发现，使流行的项目更受欢迎。为了解决这个问题，[1,11,49]引入了正则化器来引导模型给出更均衡的结果;[50,54]通过因果推理来解决人气偏差，从而解开用户兴趣和物品人气的影响。(2)系统地、不公平地歧视某些个人或群体，以有利于他人，产生所谓的不公平。由于不同的用户(或物品)组在数据中的表现通常是不平等的，因此会发生不公平。当对这种不平衡的数据进行训练时，模型很可能学习这些代表性过高的群体，并可能歧视其他代表性不足的群体[13,43]。有四种处理不公平的策略，包括重新平衡[41]、规则化[51]、对抗性学习[4]、因果推理[26]。我们鼓励读者参考[5]调查以了解更多细节。

推荐中的元学习。元学习是一种自动学习算法，旨在利用元数据来提高现有学习算法的性能或学习(诱导)学习算法本身[14,20,46]。RS中也有一些引入元学习的工作，例如[10,36]利用元学习来学习更细粒度的正则化参数;[53]利用元学习来指导推荐模型的再培训，以获得更好的绩效;元学习也被用来解决冷启动问题[30]。

1 INTRODUCTION
Being able to provide personalized suggestions to each user, rec-
ommender systems (RS) have been widely used in countless online
applications. Recent years have witnessed flourishing publications
on recommendation, most of which aim at inventing machine learn-
ing model to fit user behavior data [17,44,52]. However, these
models may be deteriorated in real-world RS, as the behavior data
is often full of biases. In practice, the data is observational rather
than experimental, and is often affected by many factors, including
but not limited to self-selection of the user (selection bias) [19,33],
exposure mechanism of the system (exposure bias) [34,47], public
opinions (conformity bias) [25,29] and the display position (position
bias) [22,23]. These biases make the data deviate from reflecting
user true preference. Hence, blindly fitting data without consider-
ing the data biases would yield unexpected results,e.g.,amplifying
the long-tail effect [1] and previous-model bias [28].

推荐系统(RS)能够为每个用户提供个性化的建议，在无数的在线应用中得到了广泛的应用。近年来，推荐方面的出版物非常多，大部分都是关于发明机器学习模型来拟合用户行为数据[17,44,52]。然而，这些模型在真实的RS中可能会恶化，因为行为数据往往充满了偏差。在实践中，数据是观察性的而非实验性的，往往受到多种因素的影响，包括但不限于用户的自我选择(选择偏差)[19,33]、系统的暴露机制(暴露偏差)[34,47]、公众意见(从众偏差)[25,29]和显示位置(位置偏差)[22,23]。这些偏差使得数据无法反映用户的真实偏好。因此，盲目拟合数据而不考虑数据偏差会产生意想不到的结果，例如:，放大了长尾效应[1]和先前模型偏差[28]。

Given the wide existence of data biases and their large impact on
the learned model, we cannot emphasize too much the importance
of properly debiasing for practical RS. Existing efforts on recom-
mendation (or learning-to-rank) biases can be divided into three
major categories: 1) data imputation [19,42], which assigns pseudo-
labels for missing data to reduce variance, 2) inverse propensity
scoring (IPS) [39,47], a counterfactual technique that reweighs the
collected data for an expectation-unbiased learning, and 3) genera-
tive modeling [27], which assumes the generation process of data
and reduces the biases accordingly. Despite their effectiveness in
some scenarios, we argue that they suffer from two limitations:

考虑到数据偏差的广泛存在及其对学习模型的巨大影响，我们再怎么强调适当去偏对于实际RS的重要性也不为过。现有的关于推荐(或学习排序)偏差的努力可以分为三大类:1)数据归责(19日42),这对缺失的数据分配pseudolabels减少方差,2)逆倾向得分(IPS)(39岁,47),一种反事实的技术,reweighs expectation-unbiased学习收集到的数据,和3)生成建模[27],假设数据的生成过程,并相应地减少偏见。尽管它们在某些情况下是有效的，但我们认为它们有两个局限性:

Lacking Universality. These methods are designed for address-
ing one or two biases of a specific scenario,e.g.,IPS for selection
bias [39], click model for position bias [12]. Thus, when facing
real data that commonly contains multiple types of biases, these
methods will fall short.

缺乏普遍性。这些方法是为解决特定场景中的一两个偏差而设计的，例如:，IPS为选择偏差[39]，点击模型为位置偏差[12]。因此，当面对通常包含多种偏见的真实数据时，这些方法将是不够的。

Lacking Adaptivity. The effectiveness of these methods is guar-
anteed only when the debiasing configurations (e.g.,pseudo-
labels, propensity scores, or data-generating process) are prop-
erly specified. However, obtaining such proper configurations is
quite difficult, requiring domain expertise that thoroughly un-
derstands the biases in the data and how they affect the model.
Even worse, the optimal configurations may evolve with time as
arXiv:2105.04170v4 [cs.LG] 8 Sep 2021
new users/items/interactions may change the data distribution,
which has been a nightmare for practitioners to manually tune
the configurations continually.

缺乏自适应性。只有在正确指定去偏配置(例如伪标签、倾向评分或数据生成过程)时，才能保证这些方法的有效性。然而，获得这样的正确配置是相当困难的，需要领域专家彻底理解数据中的偏差以及它们如何影响模型。更糟糕的是，最佳配置可能会随着时间的推移而变化，如arXiv:2105.04170v4 [cs]。新用户/项目/交互可能会改变数据分布，这对从业人员来说是一个噩梦，他们需要不断手动调整配置。

Considering the shortcomings of existing work, we believe it is
essential to develop a universal debiasing solution, which not only
accounts for multiple biases and their combinations, but also frees
human efforts to identify biases and tune the configurations. To
achieve this goal, we first review the common biases and debiasing
strategies, offering two important insights: (1) Although different
biases have different origins and properties, they all can be formu-
lated as therisk discrepancybetween the empirical risk and the true
risk, resulting from the inconsistency between the distribution for
which the training data is collected and the one used for unbiased
test; (2) The success of most recent debiasing strategies can be at-
tributed to their specific configurations to offset the discrepancy
for model training. Based on the insights, we propose a general
debiasing framework by reducing the risk discrepancy, which sub-
sumes most debiasing strategies — each strategy can be recovered
by specifying the parameters of the framework. This framework
provides a valuable opportunity to develop a universal debiasing so-
lution for recommendation — we can perform automatic debiasing
by learning the debiasing parameters of the framework.

考虑到现有工作的不足，我们认为开发一种普遍的去偏解决方案是至关重要的，它不仅能解释多种偏差及其组合，还能解放人类的努力来识别偏差和调整配置。为了实现这一目标，我们首先回顾了常见的偏见和去偏见策略，提供了两个重要的见解:(1)虽然不同的偏差有不同的起源和性质，但它们都可以表示为经验风险和真实风险之间的风险差异，这是由于收集训练数据的分布与无偏检验的分布不一致造成的;(2)最近的去偏策略的成功可以归因于它们特定的配置来抵消模型训练的差异。在此基础上，我们提出了一个降低风险差异的一般去偏框架，该框架包含了大多数去偏策略，每个策略都可以通过指定框架的参数来恢复。这个框架提供了一个宝贵的机会来开发一个通用的去偏解决方案的推荐-我们可以执行自动去偏通过学习框架的去偏参数。

Now the question lies in how to optimize the debiasing parame-
ters. Obviously, the biased training data lacks important signals of
how the data is biased and what the unbiased data looks like. To
deal with this problem, we propose to leverage anotheruniform
datato supervise the learning of debiasing parameter. The uniform
data is assumed to be collected by a random logging policy [39],
reflecting user preference in an unbiased way. We make full use
of this important evidence, optimizing the debiasing parameters
by minimizing the loss on the uniform data. Specifically, we for-
mulate the process as a bi-level optimization problem, where the
debiasing parameters serve as the hyper-parameters for learning
the recommender model, and optimize debiasing parameters by
meta-learning technique [14]. We conduct theoretical analyses on
the learning framework, proving that: (1) the optimum learned
under such objective is approximate to the best case where biases
are properly corrected; (2) it is able to learn a satisfactory debiasing
strategy even if it is trained on a small uniform data.

现在的问题是如何优化去偏参数。显然，有偏差的训练数据缺乏重要的信号，表明数据是如何有偏差的，以及无偏差的数据是什么样的。为了解决这个问题，我们提出利用另一个统一的数据来监督去偏参数的学习。假设统一数据是由随机日志记录策略[39]收集的，以无偏的方式反映用户偏好。我们充分利用这一重要证据，通过最小化均匀数据的损耗来优化去偏参数。具体地说，我们将这个过程描述为一个双层优化问题，其中去偏参数作为学习推荐模型的超参数，并通过元学习技术[14]优化去偏参数。我们对学习框架进行了理论分析，证明:(1)该目标下的最优学习近似于偏差得到正确修正的最优情况;(2)即使在小的均匀数据上训练，也能学习到满意的去偏策略。

最后，在利用统一数据进行推荐方面，最相关的工作是最近提出的KDRec[28]。然而，我们认为它没有充分利用统一数据的优点。KDRec在统一的数据上训练一个单独的教师模型，然后将模型的知识转移到有偏差的数据上的正常训练。由于统一数据的收集是以降低用户体验为代价的，它的大小通常是相当小的。因此，在此基础上训练的模型方差较大，降低了KDRec的有效性。而且缺乏理论保障，教师模式如何抵消偏差的内在机制也没有被完全理解。与KDRec相比，我们的框架以更合理的理论方式使用统一数据，并产生显著的经验改进。

In a nutshell, this work makes the following main contributions:

Unifying various biases from therisk discrepancyperspec-
tive and developing a general debiasing framework that sub-
sumes most debiasing strategies.

总而言之，本文的主要贡献如下:

从风险差异的角度统一各种偏见，并发展一个包含大多数去偏策略的总体去偏框架。

Proposing a new method that leverages uniform data to learn
optimal debiasing strategy with theoretical guarantees.

•Conducting experiments on three types of data (explicit and
implicit feedback, and simulated data of list feedback) to
validate the effectiveness of our proposal.

提出了一种利用统一数据学习最优去偏策略的新方法。

•对三种类型的数据(显式反馈、隐式反馈、列表反馈模拟数据)进行实验，验证我们的建议的有效性。

思路：rs好》有问题》其它人的解决方案》他们的方法不好，有缺陷；》我的方法好；有哪些难点，

我要做什么？看论文；

把论文下载下来，翻译，直接看；recsys2021先下载几篇关于bias的，先看；然后在下载其余的文献；一篇论文要看好久的；

csdn和知乎发布的文章能再次编辑吗？可以

现发布占领位置，然后在修改；留名字，微信，公众号；

在网上看看B站和知乎，看看人家是怎么做科研的，经验分享；

10/11论文关键词，自动去偏框架论文翻译，发文章思考相关推荐

SIGIR 2021 | AutoDebias：推荐系统中通用的自动去偏方法
©PaperWeekly 原创 · 作者|董汉德学校|中国科学技术大学硕士生研究方向|信息检索摘要近些年来,推荐系统的偏差越来越受到研究者的关注.训练推荐模型的数据大多为观测所得,而非实验所得 ...
推荐系统去偏（Debiased Recommendation）研究进展概述
©作者 | 张景森学校 | 中国人民大学信息学院硕士文章来源 | RUC AI Box 引言推荐系统作为解决信息过载的一种重要手段,已经在不同的应用场景下取得了不错的效果.近些年来关于推荐系统的 ...
2022年全球自动去鱼皮机行业分析报告
2022年全球市场自动去鱼皮机总体规模.主要生产商.主要地区.产品和应用细分研究报告本文研究全球市场.主要地区和主要国家自动去鱼皮机的销量.销售收入等,同时也重点分析全球范围内主要厂商(品牌)竞争态 ...
医学SCI论文关键词的选取方法
在一篇完整的医学SCI论文中,关键词是非常细小的一部分,但这并不影响关键词在SCI论文中所起的重要作用.不少作者在选取关键词的时候常常会感到困惑,是不是随便选取几个关键词就好了.其实对于这一点,作者无 ...
【论文写作】如何写科技论文？万能模板！！！（以IEEE会议论文为例）
0. 写在前面常言道,科技论文犹如"八股文",有固定的写作模式.本篇博客主要是针对工程方面的论文的结构以及写作链条的一些整理,并不是为了提高或者润色一篇论文的表达.基本上所有的论 ...
高等数学在计算机方面的应用论文2000字,数学生活应用论文范文
1.利用生活情境开展小学数学教学的方法在小学数学教学中应用生活情境的方法多种多样,下面就对生活情境在小学数学教学中的应用进行分析. 1.1利用生活情境,激发小学生的学习兴趣在小学数学教学中,教师只 ...
Interview：算法岗位面试—10.11下午—上海某公司算法岗位(偏数据分析，证券金融行业)技术面试考点之sqlserver语言相关考察点复习
Interview:算法岗位面试-10.11下午-上海某公司算法岗位(偏数据分析,证券金融行业)技术面试考点之sqlserver语言相关考察点复习导读:其实,考察的知识点,博主都做过, 还包括sql ...
Interview：算法岗位面试—10.11下午—上海某公司算法岗位(偏机器学习，互联网数字行业)技术面试考点之XGBoost的特点、python的可变不可变的数据类型、赋值浅拷贝深拷贝区别
ML岗位面试:10.11下午-上海某公司算法岗位(偏机器学习,互联网数字行业)技术面试考点之XGBoost的特点.python的可变不可变的数据类型.赋值浅拷贝深拷贝区别 Interview:算法岗位 ...
ACL 2019开源论文 | 句对匹配任务中的样本选择偏差与去偏方法
作者丨张冠华单位丨腾讯实习生 & 哈工大本科生研究方向丨自然语言处理导读句对匹配 (Natural Language Sentence Matching,NLSM) 任务是指给定两个句 ...

10/11论文关键词，自动去偏框架论文翻译，发文章思考

10/11论文关键词，自动去偏框架论文翻译，发文章思考相关推荐

最新文章

热门文章