晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

共形群组推荐系统:给「一群人」的推荐配上置信度

准大一 · 计算机科学与技术 × 推荐系统 × 可信机器学习 —— 群组推荐置信度精读材料
原文:arXiv:2307.12034 2023年7月22日发布 arXiv 预印本(cs.IR) 群组推荐 × 共形预测 × 置信度 附英文摘要朗读音频

一、论文档案

英文标题Conformal Group Recommender System
中文标题共形群组推荐系统(CGRS):把共形预测引入群组推荐,让每次推荐都附带置信度与错误上界
作者文卡特斯瓦拉·拉奥·卡吉塔, 安舒曼·辛格, 维卡斯·库马尔, 帕万·卡尔扬·雷迪·尼鲁杜, 阿伦·K·普贾里, 罗希特·库马尔·邦杜古拉(机构未在素材中标注)
发布时间2023年7月22日(v1)|分类:cs.IR(信息检索)
一句话概括传统群组推荐是「黑盒」——只告诉一群人推荐什么,不告诉有多靠谱;这篇论文用共形预测给推荐集附上「置信度 (1−ε)」,系统说 80% 置信度,犯错概率就至多 20%。
💡 为什么选这篇给你:① 推荐系统是你每天在用(刷视频、点外卖、订酒店)的技术,群组推荐更是「全家一起看电影」场景的刚需;② 思路清爽——不堆模型,而是把一个通用框架(共形预测)嫁接到群组推荐上,理论性质(可交换性、有效性、效率)讲得清清楚楚;③ 数字直观:置信度 80% ⇔ 错误率 ≤ 20%,这种「可承诺的可靠性」正是推荐系统研究的新方向。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
recommender system (RS)推荐系统帮用户从海量商品(电影、网页、图书)里筛出相关物品的系统,靠历史交互做个性化推荐。
group recommender system (GRS)群组推荐系统给一群人(家人、朋友、同事)推荐同一个物品——一起看电影、聚餐、出游。
content-based filtering基于内容的过滤按「物品内容/特征与用户画像的相似度」推荐。
collaborative filtering协同过滤按「和你有相似偏好的其他用户」的喜好推荐。
hybrid filtering混合过滤组合多种机制与多种数据源的推荐方式。
conformal prediction共形预测一个可靠机器学习框架:对每个样本给出「预测 + 置信度」,可配合任意传统算法使用。
nonconformity measure非一致性度量打分「新样本和历史样本有多不合群」,是共形框架的核心零件。
significance level ε显著性水平允许的犯错概率上限;推荐集的置信度 = 1−ε。
exchangeability可交换性样本顺序不影响联合分布——共形预测理论成立的前提性质。
validity有效性实际错误率真的被 ε 界住(不「欠覆盖」、不吹牛)。
efficiency效率在保证有效性的前提下,预测/推荐集尽量小、尽量精准。
confidence measure置信度度量系统对「期望的物品在推荐集里」这件事有多自信的量化。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
Group recommender systems (GRS) are critical in discovering relevant items from a near-infinite inventory based on group preferences rather than individual preferences, like recommending a movie, restaurant, or tourist destination to a group of individuals.
CN · 翻译
群组推荐系统(GRS)至关重要——它基于群体偏好而非个人偏好,从近乎无限的候选项中发现相关物品,比如给一群人推荐电影、餐厅或旅游目的地。
EN · 原文
The traditional models of group recommendation are designed to act like a black box with a strict focus on improving recommendation accuracy, and most often, they place the onus on the users to interpret recommendations.
CN · 翻译
传统的群组推荐模型被设计成黑盒,一心只追求提升推荐准确率,而且常常把「理解推荐结果」的责任推给用户
EN · 原文
In recent years, the focus of Recommender Systems (RS) research has shifted away from merely improving recommendation accuracy towards value additions such as confidence and explanation.
CN · 翻译
近年来,推荐系统(RS)研究的重心已从「单纯提升准确率」转向置信度、可解释性等价值增量。
EN · 原文
In this work, we propose a conformal prediction framework that provides a measure of confidence with prediction in conjunction with a group recommender system to augment the system-generated plain recommendations.
CN · 翻译
本文提出一个共形预测框架:与群组推荐系统配合,为每次预测附带一个置信度度量,增强系统生成的「裸推荐」。
EN · 原文
In the context of group recommender systems, we propose various nonconformity measures that play a vital role in the efficiency of the conformal framework.
CN · 翻译
针对群组推荐场景,我们提出多种非一致性度量——它们在共形框架的效率中扮演关键角色。
EN · 原文
We also show that defined nonconformity satisfies the exchangeability property.
CN · 翻译
我们还证明了所定义的非一致性度量满足可交换性性质。
EN · 原文
Experimental results demonstrate the effectiveness of the proposed approach over several benchmark datasets.
CN · 翻译
实验结果表明,所提方法在多个基准数据集上都是有效的。
EN · 原文
Furthermore, our proposed approach also satisfies validity and efficiency properties.
CN · 翻译
此外,我们的方法还满足有效性(validity)与效率(efficiency)两项性质。

关键词 Keywords:Group Recommender System 群组推荐系统 | Conformal Prediction 共形预测 | Nonconformity Measure 非一致性度量 | Confidence 置信度

四、引言精选(为什么这个问题重要)

① 从「个性化推荐」到「群组推荐」:给一家人推荐同一部电影

EN · 原文
Recommender systems (RS) assist users in decision-making by helping them sift through a huge variety of offered products, such as movies, web pages, articles, and books [1]. These systems exploit past interactions between users and items to make personalized recommendations. RS algorithms are broadly classified into content-based, collaborative, and hybrid filtering, depending on the input used for profiling users and items. Content-based filtering approaches recommend items to a user by considering the similarity between the content or features of items and the user's profile [2, 3]. On the other hand, the collaborative filtering-based approach recommends items based on the preferences of other users who share similar preferences to the target user [4, 5, 6]. The hybrid approach combines various mechanisms and compositions of different data sources [7, 8].
CN · 翻译
推荐系统(RS)帮用户做决策——在电影、网页、文章、图书等海量商品中筛选。它们利用用户与物品的历史交互做个性化推荐,按输入可分为三类:基于内容(看物品内容/特征与用户画像的相似度)、协同过滤(看相似偏好用户的喜好)、混合(组合多种机制与数据源)。
EN · 原文
Group recommender systems (GRS) [9, 10] extend this concept to the user group, wherein it analyses a group of users' profiles and creates a communal recommendation list. These groups can consist of family members, friends, colleagues, or any collection of individuals who wish to engage with a specific item or application collectively. We observe numerous applications of group recommender systems in our daily lives. For example, a group of friends may plan to dine at a restaurant or organize a tour, while a family may wish to watch a movie together.
CN · 翻译
群组推荐系统(GRS)把这个概念扩展到用户群体:分析一组用户的画像,生成一份「集体推荐清单」。群体可以是家人、朋友、同事,或任何想共同参与某个物品/应用的个体集合。日常生活里到处是它的身影——一群朋友商量去哪家餐厅、组织一次旅行,一家人想一起看部电影。

② 从「只追准确率」到「置信度与可解释性」:黑盒推荐的信任危机

EN · 原文
Several approaches in recent years have been proposed to extend personalized recommender systems to group recommender systems, wherein the major focus is to improve the accuracy of recommendations. With the recent algorithmic advancement, the focus of the research has been recently shifted towards creating transparent group recommendation models that prioritize accountability and explainability. These models aim to achieve accuracy while providing additional value through confidence measures, explanations, or sensitivity. Among these enhancements, associating a confidence measure with the recommendation set is a particularly appealing facet. The confidence measures indicate the system's confidence that the desirable items are present in the recommendation set, which in turn enhance the system's reliability and help users quickly and correctly identify the products of their choice. Although researchers have thoroughly investigated confidence-based personalized recommender systems [11, 12], only a few methods have been proposed for improving the confidence measure in group recommender systems. Further, considering the individual preferences of the group and providing a delightful recommendation with confidence is not trivial.
CN · 翻译
近年来许多工作把个性化推荐扩展到群组推荐,但主战场仍是提升准确率。随着算法进步,研究重心转向透明、可问责、可解释的群组推荐模型——在保证准确率的同时,通过置信度、解释或敏感性提供额外价值。其中,给推荐集附上置信度度量尤其吸引人:它表明系统对「期望物品在推荐集中」有多自信,能增强可靠性,帮用户快速、正确地找到心仪商品。虽然置信度个性化推荐已被充分研究,但群组推荐上的置信度方法寥寥无几——何况要兼顾群体里每个人的偏好、还给出可信推荐,绝非易事。
EN · 原文
This paper proposes a confidence-based group recommender system using a conformal prediction framework. The proposed approach represents confidence in connection with the error bound, i.e., if the system exhibits 80% confidence in the recommendation, it implies that the probability of making an error is at most 20%. The concept of conformal prediction forms the basis for the proposed Conformal Group Recommender System (CGRS).
CN · 翻译
本文提出一个基于共形预测框架的置信度群组推荐系统。置信度与错误上界直接挂钩:如果系统对推荐给出 80% 的置信度,就意味着犯错概率至多 20%。共形预测构成了所提出的共形群组推荐系统(CGRS)的基础。

③ 共形预测:给「推荐」配上一个错误上界

EN · 原文
Conformal prediction is a framework for reliable machine learning that measures confidence in the predicted labels on a per-instance basis. A conformal system framework can be implemented alongside any traditional machine learning algorithm, but the implementation largely depends on the underlying algorithm. This paper extends the concept to GRS and defines a nonconformity measure, an essential part of the conformal prediction framework, suitable for group recommendation settings. Given a set of users U, items O, and a target group of user profiles, Oj (∀j), UG and the significance level ε, the proposed conformal group recommender system returns recommendation set to a group UG with the confidence of (1−ε). We demonstrate that the exchangeability property, an essential property for any conformal framework, is satisfied by the proposed nonconformity measure. Furthermore, the proposed approach also satisfies the validity and efficiency properties. Experiment results over several real-world datasets support the efficacy of the proposed conformal group recommender system.
CN · 翻译
共形预测是一个可靠机器学习框架,按样本逐例地度量预测标签的置信度,可与任意传统机器学习算法搭配实现。本文把它扩展到群组推荐,定义了适合群组场景的非一致性度量(共形框架的核心部件)。给定用户集 U、物品集 O、目标群体的用户画像以及显著性水平 ε,所提系统向群体 UG 返回置信度为 (1−ε) 的推荐集。我们证明了该非一致性度量满足共形框架成立的关键性质——可交换性;同时满足有效性效率性质,多个真实数据集上的实验支持了其有效性。
💡 这是全文最有味道的一句“if the system exhibits 80% confidence in the recommendation, it implies that the probability of making an error is at most 20%.”——推荐系统第一次给出「可验证的承诺」:置信度不是玄学,而是严格绑定的错误上界。把「推荐」从猜测变成「有边界的承诺」,这就是共形预测的价值。

五、论文贡献(3 个要点)

EN · 原文
1. In this work, we propose a conformal prediction framework that provides a measure of confidence with prediction in conjunction with a group recommender system to augment the system-generated plain recommendations.
CN · 翻译
1. 带置信度的群组推荐共形框架。提出与群组推荐系统结合的共形预测框架:为预测附带置信度度量,把系统生成的「裸推荐」升级为「可信推荐」。
EN · 原文
2. In the context of group recommender systems, we propose various nonconformity measures that play a vital role in the efficiency of the conformal framework. We also show that defined nonconformity satisfies the exchangeability property.
CN · 翻译
2. 多种非一致性度量 + 可交换性证明。针对群组推荐提出多种非一致性度量(决定共形框架的效率),并证明其满足共形框架成立的前提——可交换性
EN · 原文
3. The proposed approach represents confidence in connection with the error bound, i.e., if the system exhibits 80% confidence in the recommendation, it implies that the probability of making an error is at most 20%. Furthermore, our proposed approach also satisfies validity and efficiency properties. Experimental results demonstrate the effectiveness of the proposed approach over several benchmark datasets.
CN · 翻译
3. 置信度绑定错误上界,有效且高效。置信度与错误上界严格挂钩(80% 置信度 ⇒ 犯错概率 ≤ 20%),同时满足有效性与效率性质,并在多个基准数据集上验证了有效性。

六、结论中英对照

EN · 原文
This paper introduces a conformal framework to the group recommendation scenario for a reliable recommendation. The theoretical facets in the article demonstrate the likelihood that the proposed CGRS makes an error is bounded by the given significance level ε, and hence the system exhibits a confidence of (1−ε).
CN · 翻译
本文把共形框架引入群组推荐场景,以实现可靠的推荐。理论部分证明:所提出的 CGRS 出错的概率被给定的显著性水平 ε 界住,因此系统呈现 (1−ε) 的置信度。
EN · 原文
In addition to furnishing a confidence measure of reliability, the proposed method also improves the quality of recommendations. Our experimental analysis of various benchmark datasets corroborates that the proposed CGRS performs better than the baseline GRS approach in terms of different standard performance metrics assessing recommendation quality.
CN · 翻译
除了提供可靠性置信度度量,该方法还提升了推荐质量。在多个基准数据集上的实验分析证实:按评估推荐质量的多项标准指标衡量,CGRS 都优于基线 GRS 方法。
EN · 原文
Extension of the proposed framework to various group recommendation algorithms is a goal worth pursuing in the future. Further, investigating a conformal framework that efficiently furnishes confidence to the complex group recommender algorithms, such as deep learning-based models, is also an exciting direction for vigorous research.
CN · 翻译
把该框架扩展到各种群组推荐算法是未来值得追求的目标;此外,为深度学习等复杂群组推荐算法高效地提供置信度的共形框架,也是一个激动人心的研究方向。

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:给一个人推荐已经很难,给「一群人」推荐更难——而且传统模型是黑盒:只甩给你一个推荐列表,不告诉你「它到底靠不靠谱」。电影选错了全家扫兴,谁来负责?
  2. 做法:把共形预测(conformal prediction)这个「带错误上界的机器学习框架」装进群组推荐。核心是设计适合群组的非一致性度量(衡量一个物品跟这个群体的历史偏好有多「不合群」),据此构造带置信度 (1−ε) 的推荐集,并严格证明可交换性、有效性、效率三条性质。
  3. 结果:在多个基准数据集上,CGRS 不仅给出可信的置信度,推荐质量还比基线 GRS 更好——可靠性不是用牺牲质量换来的。
  4. 最值钱的观点:推荐系统研究的重心正从「准确率」转向「置信度与可解释性」;「80% 置信度 = 犯错概率 ≤ 20%」这种可验证的承诺,比一个无法解释的高准确率更有价值。
  5. 工程意义:对电商、流媒体、旅游平台来说,群体推荐(家庭套餐、团建方案)是真实场景;置信度让用户敢用、让系统可问责,也为「推荐失败谁来背锅」提供了数学依据。
🎯 对保研的启示:这篇论文示范了「老问题 + 新框架」的选题套路——群组推荐是老问题,共形预测是成熟框架,组合起来就是一篇理论扎实、实验完整的工作。复试时讲「我如何发现两个领域的结合点、如何证明新方法的性质」,比堆砌模型名词更打动导师。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么(群组推荐不可信)?方法是什么(共形预测)?结果是什么(有效 + 高效)?
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「从准确率到置信度」这个研究转向,以及 (1−ε) 置信度承诺为什么是「可验证的」。
  3. 第三遍(30 分钟):读方法文字部分(非一致性度量、可交换性、有效性/效率),跳过所有公式和证明,只看文字描述;遇到不懂的术语回查术语表。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出每个数字(80% confidence、20% error)和术语(conformal prediction、nonconformity measure、exchangeability)。