围观RecSys2020 | 推荐系统顶会说了啥？(附论文打包下载)

---

RecSys2020主页：RecSys – ACM Recommender Systems(https://recsys.acm.org/)

本仙女整理了RecSys2020的文章，其中包括41篇Long Papers，26篇Short Papers和10篇Industry Papers。

---

PART ONE 客观解读

本文整理了RecSys2020的文章，其中包括41篇Long Papers，26篇Short Papers和10篇Industry Papers。基于个人阅读习惯抓取了几个思路进行总结。

1. 今年的RecSys有相当数量的文章关注了Bandits方法以及强化学习的思路。

• Cascading Hybrid Bandits：Online Learning to Rank for Relevance and Diversity

• Contextual User Browsing Bandits for Large-Scale Online Mobile Recommendation

• Exploring Clustering of Bandits for Online Recommendation System

• Global and Local Differential Privacy for Collaborative Bandits

• Offline Contextual Multi-armed Bandits for Mobile Health Interventions：A Case Study on Emotion Regulation

• Carousel Personalization in Music Streaming Apps with Contextual Bandits

• Contextual Meta-Bandit for Recommender Systems Selection

• Deep Bayesian Bandits：Exploring in Online Personalized Recommendations

• Keeping Dataset Biases out of the Simulation : A Debiased Simulator for Reinforcement Learning based Recommender Systems

• Learning to Collaborate in Multi-Module Recommendation via Multi-Agent Reinforcement Learning without Communication

• Content-Collaborative Disentanglement Representation Learning for Enhanced Recommendation.

在之前，Bandits算法簇就已经广泛应用在推荐系统中，但是大多数情况下用来处理冷启动问题，例如汤普森采样等。在今年的会议中，我们可以看到，大家已经将Bandits的指针指向了在线学习、上下文环境以及贝叶斯方法。此外，这其中还有一篇文章将Bandits方法与隐私问题相结合。

此外，也有强化学习的加入。使用强化学习的博弈的思想可以很好的解决推荐系统中的偏差问题。

2. 今年的RecSys更关注多样的应用场景。

• 一直倍受关注的评分与评论问题：Rating and Review Data by Initializing Latent Factor Models with Topic Models for Top-N Recommendation

• 音乐推荐场景：Contextual and Sequential User Embeddings for Large-Scale Music Recommendation ；Carousel Personalization in Music Streaming Apps with Contextual Bandits

• 新闻推荐场景：KRED:Knowledge-Aware Document Representation for News Recommendations

• 信号预测场景：Making Neural Networks Interpretable with Attribution: Application to Implicit Signals Prediction

• 视频推荐场景：Recommending the Video to Watch Next: An Offline and Online Evaluation at YOUTV.de

此外，还有图书、座位推荐、在线游戏场景，甚至还有财务、家庭能源浪费、拍卖、分析师、马拉松运动员等新奇的应用方向。可以看到，推荐系统领域的个性化，不仅仅体现在每一个用户的个性化偏好，也体现在个性化的应用领域。

3. 今年的RecSys更关注隐私与安全。

获取用户记录来进行用户偏好建模这一问题，其实确实存在一定的隐私和安全问题。该问题值得在推荐系统顶会中被广泛关注。

隐私：Global and Local Differential Privacy for Collaborative Bandits

攻击：Revisiting Adversarially Learned Injection Attacks Against Recommender Systems

安全：Towards Safety and Sustainability_ Designing Local Recommendations for Post-pandemic World

4. 上下文话题也频繁出现。

• 上下文的Bandits问题：Contextual User Browsing Bandits for Large-Scale Online Mobile Recommendation ；Carousel Personalization in Music Streaming Apps with Contextual Bandits ；Contextual Meta-Bandit for Recommender Systems Selection

• 上下文的用户嵌入问题：Contextual and Sequential User Embeddings for Large-Scale Music Recommendation

• 结合注意力的上下文问题：TAFA: Two-headed Attention Fused Autoencoder for Context-Aware Recommendations

• 游戏场景中的上下文：Interpretable Contextual Team-aware Item Recommendation: Application in Multiplayer Online Battle Arena Games

• 查询与上下文：Query as Context for Item-to-Item Recommendation

5. Transformer以及BERT仍旧没有退出RecSys。

但是使用Transformer以及BERT的目的，和NLP领域有所不同。在推荐场景中， 更关注它们能够学到什么用户偏好信息，这与文本的语义信息是不同的。

• SSE-PT:Sequential Recommendation Via Personalized Transformer

• What does BERT know about books, movies and music：Probing BERT for Conversational Recommendation

6. 协同过滤是永远的神。

不解释，问就是都得读。

• Neural Collaborative Filtering vs. Matrix Factorization Revisited

• Theoretical Modeling of the Iterative Properties of User Discovery in a Collaborative Filtering Recommender System

• ClusterExplorer：Enable User Control over Related Recommendations via Collaborative Filtering and Clustering

• Explainable Recommendations via Attentive Multi-Persona Collaborative Filtering

• History-Augmented Collaborative Filtering for Financial Recommendations

7. 工业应用文章的崛起。

今年的RecSys收录了十篇Industry Papers。可以看到，这些文章的立意与Motivation是与学术界探索的话题有所不同的。这些文章大多数立足于具体的工业应用场景，例如Amazon视频、宜家APP、约会APP以及学术职业生涯平台等。

• Balancing Relevance and Discovery to Inspire Customers in the IKEA App

• Behavior-based Popularity Ranking on Amazon Video

• Building a reciprocal recommendation system at scale from scratch：Learnings from one of Japan's prominent dating applications

• Developing Recommendation System to provide a Personalized Learning experience at Chegg

• Investigating Multimodal Features for Video Recommendations at Globoplay

• On the Heterogeneous Information Needs in the Job Domain： A Unified Platform for Student Career

PART TWO 个人观点

推荐系统千人千面，这推荐系统顶会文章每一个人读起来，感触也会不同。以下几点是我个人觉得本次会议中较为有趣的点，分享给大家。

• Point1 ：Debiasing Item-to-Item Recommendations With Small Annotated Datasets

这篇文章关注了小数据集。小样本学习近两年来非常火热，个人觉得也非常的有意义。尤其是在推荐系统中，由于用户数据常常伴随着极大的稀疏性，如何利用少量的用户记录来学习到更好的用户偏好是一个非常重要的问题。本次会议中关于小样本的推荐系统学习论文不多，目前仅看到这一篇。推荐阅读。

• Point2 ：ImRec:Learning Reciprocal Preferences Using Images。

这似乎是唯一一篇关注了图像的论文。推荐系统中图像也应该有较高的关注，例如在购物网站中。近年来多模态学习越来越变成了研究热潮，或许多模态学习能够进一步地推动图像在推荐系统中的应用。期待明年能有更多的多模态推荐文章。

---

PART THREE 论文列表

为了便于大家快速浏览，对于每一篇paper，本文给出了一个粗糙的标题翻译。

Long Papers

• 一种将业务指标匿名发布隐式反馈数据集的方法 | A Method to Anonymize Business Metrics to Publishing Implicit Feedback Datasets

• 会话推荐系统潜在线性评论的一种排序优化方法 | A Ranking Optimization Approach to Latent Linear Critiquing for Conversational Recommender Systems

• 我们是在评估对可重复评估和公平比较的严格的基准化建议吗 | Are We Evaluating Rigorously：Benchmarking Recommendation for Reproducible Evaluation and Fair Comparison

• 级联的混合Bandits：在线学习的相关性和多样性排名 | Cascading Hybrid Bandits：Online Learning to Rank for Relevance and Diversity

• 增强推荐的内容协同解解耦表示学习 | Content-Collaborative Disentanglement Representation Learning for Enhanced Recommendation.

• 大规模音乐推荐的上下文和顺序用户嵌入 | Contextual and Sequential User Embeddings for Large-Scale Music Recommendation

• 上下文用户浏览bandits大规模在线移动推荐 | Contextual User Browsing Bandits for Large-Scale Online Mobile Recommendation

• 使用小型带注释的数据集消除物品间推荐的偏差 | Debiasing Item-to-Item Recommendations With Small Annotated Datasets

• 解构过滤气泡:用户决策和推荐系统 | Deconstructing the Filter Bubble: User Decision-Making and Recommender Systems.

• 双重鲁棒估计与点击后转换的排名指标 | Doubly Robust Estimator for Ranking Metrics with Post-Click Conversions

• 通过对排名敏感的相关性平衡，确保群组推荐的公平性 | Ensuring Fairness in Group Recommendations by Rank-Sensitive Balancing of Relevance

• 利用性能估计来增强推荐集合 | Exploiting Performance Estimates for Augmenting Recommendation Ensembles.

• 探索在线推荐系统的Bandits聚类 | Exploring Clustering of Bandits for Online Recommendation System

• 将项目相似度模型与自注意网络融合，进行顺序推荐 | FISSA:Fusing Item Similarity Models with Self-Attention Networks for Sequential Recommendation

• 从实验室到生产:家庭改善领域基于会话的推荐案例研究 | From the lab to production: A case study of session-based recommendations in the home-improvement domain

• 全局和局部不同的隐私为协同Bandits | Global and Local Differential Privacy for Collaborative Bandits

• 针对分析师的目标驱动命令推荐 | Goal-driven Command Recommendations for Analysts

• ImRec:学习互惠偏好使用图像 | ImRec:Learning Reciprocal Preferences Using Images

• 店内增强现实的产品比较和推荐 | In-Store Augmented Reality-Enabled Product Comparison and Recommendation

• 避免数据集的偏差在模拟:一个去偏的模拟器为强化学习基于推荐系统 | Keeping Dataset Biases out of the Simulation : A Debiased Simulator for Reinforcement Learning based Recommender Systems

• KRED:用于新闻推荐的知识感知文档表示 | KRED:Knowledge-Aware Document Representation for News Recommendations

• 通过不需要交流的多智能体强化学习，学习在多模块推荐中协同 | Learning to Collaborate in Multi-Module Recommendation via Multi-Agent Reinforcement Learning without Communication

• 用属性使神经网络可解释:隐式信号预测的应用 | Making Neural Networks Interpretable with Attribution: Application to Implicit Signals Prediction

• MultiRec:拍卖系统中唯一物品推荐的一种多关系方法 | MultiRec: A Multi-Relational Approach for Unique Item Recommendation in Auction Systems

• 重新讨论神经协同过滤与矩阵分解 | Neural Collaborative Filtering vs. Matrix Factorization Revisited

• 线下背景多臂Bandits对移动健康的干预——以情绪调节为例 | Offline Contextual Multi-armed Bandits for Mobile Health Interventions：A Case Study on Emotion Regulation

• 离线推荐系统评价中的目标物品抽样问题 | On Target Item Sampling in Offline Recommender System Evaluation

• 渐进式分层抽取(PLE)_一种新的个性化推荐多任务学习(MTL)模型 | Progressive Layered Extraction (PLE)_ A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations

• PURS:个性化意外推荐系统，提高用户满意度 | PURS: Personalized Unexpected Recommender System for Improving User Satisfaction

• 作为图探索的推荐 | Recommendations as Graph Explorations

• 推荐接下来观看的视频:在YOUTV.de进行线下和线上评估 | Recommending the Video to Watch Next: An Offline and Online Evaluation at YOUTV.de

• RecSeats:一个混合卷积神经网络选择模型，用于预订座位地点的座位推荐 | RecSeats: A Hybrid Convolutional Neural Network Choice Model for Seat Recommendations at Reserved Seating Venues

• 重新审视对推荐系统的对抗性注入攻击 | Revisiting Adversarially Learned Injection Attacks Against Recommender Systems

• SSE-PT:通过个性化Transformer的顺序推荐 | SSE-PT:Sequential Recommendation Via Personalized Transformer

• TAFA:双头注意力融合自动编码器，用于上下文感知推荐 | TAFA: Two-headed Attention Fused Autoencoder for Context-Aware Recommendations

• 协同过滤推荐系统中用户发现迭代特性的理论建模 | Theoretical Modeling of the Iterative Properties of User Discovery in a Collaborative Filtering Recommender System

• 朝向安全和可持续发展：为大流行后的世界制定地方推荐 | Towards Safety and Sustainability_ Designing Local Recommendations for Post-pandemic World

• 无偏的广告点击预测位置感知广告系统 | Unbiased Ad Click Prediction for Position-aware Advertising Systems

• 无偏学习推荐的因果效应 | Unbiased Learning for the Causal Effect of Recommendation

• 关于书、电影和音乐，BERT 知道些什么 | What does BERT know about books, movies and music：Probing BERT for Conversational Recommendation

• “谁不喜欢恐龙”:发现并引出更丰富的推荐偏好 | "Who doesn't like dinosaurs": Finding and Eliciting Richer Preferences for Recommendation

Short Papers

• 自适应点对学习-排名基于内容的个性化推荐 | Adaptive Pointwise-Pairwise Learning-to-Rank for Content-based Personalized Recommendation

• ADER:自适应地提炼出回放范例，以实现基于会话的推荐的持续学习 | ADER：Adaptively Distilled Exemplar Replay Towards Continual Learning for Session-based Recommendation

• 具有上下文bandits的音乐流应用中的轮播个性化 | Carousel Personalization in Music Streaming Apps with Contextual Bandits

• 推荐系统的因果推理 | Causal Inference for Recommender Systems

• 允许用户通过协同过滤和聚类来控制相关推荐 | ClusterExplorer_ Enable User Control over Related Recommendations via Collaborative Filtering and Clustering

• 将潜在因素模型与主题模型初始化，结合评分和评论数据进行Top-N推荐 | Combining Rating and Review Data by Initializing Latent Factor Models with Topic Models for Top-N Recommendation

• 上下文的元bandits推荐系统选择 | Contextual Meta-Bandit for Recommender Systems Selection

• 从响应率偏差中消除用户满意度估计 | Deconfounding User Satisfaction Estimation from Response Rate Bias

• 深度贝叶斯bandits:探索在线个性化推荐 | Deep Bayesian Bandits：Exploring in Online Personalized Recommendations

• 重复使用的可解释的推荐 | Explainable Recommendation for Repeat Consumption

• 可解释的推荐，通过关注的多人物协同过滤 | Explainable Recommendations via Attentive Multi-Persona Collaborative Filtering

• 探索基于会话的推荐的纵向效应 | Exploring Longitudinal Effects of Session-based Recommendations

• 适合跑步:马拉松训练的个性化推荐 | Fit to Run：Personalised Recommendations for Marathon Training

• 免费午餐！在ROI约束内进行动态促销推荐的回顾性提升模型 | Free Lunch! Retrospective Uplift Modeling for Dynamic Promotions Recommendation within ROI Constraints

• 财务推荐的历史增强型协同过滤 | History-Augmented Collaborative Filtering for Financial Recommendations

• 在不同偏好强度下改进多任务的一类推荐 | Improving One-class Recommendation with Multi-tasking on Various Preference Intensities

• 可解释的上下文团队意识项目推荐：在多人在线战场游戏中的应用 | Interpretable Contextual Team-aware Item Recommendation：Application in Multiplayer Online Battle Arena Games

• 基于长尾会话的推荐 | Long-tail Session-based Recommendation

• 混合注意机制与多时间嵌入顺序推荐 | MEANTIME：Mixture of Attention Mechanisms with Multi-temporal Embeddings for Sequential Recommendation

• 基于频率的双哈希方法减少推荐系统的模型尺寸 | Model Size Reduction Using Frequency Based Double Hashing for Recommender Systems

• 双曲几何模型在Top-N推荐任务上的性能 | Performance of Hyperbolic Geometry Models on Top-N Recommendation Tasks

• 音乐推荐算法的个性化偏见 | Personality Bias of Music Recommendation Algorithms

• 为马拉松运动员提供可解释的比赛时间预测和训练计划建议 | Providing Explainable Race-Time Predictions and Training Plan Recommendations to Marathon Runners

• 通过实时推荐减少家庭能源浪费 | Reducing energy waste in households through real-time recommendations

• 基于组合联合学习的无偏内隐推荐和倾向估计 | Unbiased Implicit Recommendation and Propensity Estimation via Combinational Joint Learning

• 用概念上的不协调作为提出推荐的基础 | Using conceptual incongruity as a basis for making recommendations

Industry Papers

• 人类对算法相似性的看法 | A Human Perspective on Algorithmic Similarity

• 平衡相关性和发现，在宜家应用中激发顾客 | Balancing Relevance and Discovery to Inspire Customers in the IKEA App

• 基于行为的亚马逊视频人气排名 | Behavior-based Popularity Ranking on Amazon Video

• 从日本一个著名的约会应用程序中学习，建立一个规模互惠的推荐系统 | Building a reciprocal recommendation system at scale from scratch_ Learnings from one of Japan's prominent dating applications

• 反事实学习的推荐系统 | Counterfactual learning for recommender system

• 开发推荐系统，为Chegg提供个性化的学习体验 | Developing Recommendation System to provide a Personalized Learning experience at Chegg

• 为Globoplay的视频推荐研究多模式特性 | Investigating Multimodal Features for Video Recommendations at Globoplay

• 关于工作领域的异构信息需求——一个统一的学生职业生涯平台 | On the Heterogeneous Information Needs in the Job Domain_ A Unified Platform for Student Career

• 查询作为Item到Item推荐的上下文 | Query as Context for Item-to-Item Recommendation

• 来自cold的嵌入:用基于内容的推断改进新和稀有产品的载体 | The Embeddings That Came in From the Cold：Improving Vectors for New and Rare Products with Content-Based Inference

---

所有文章的pdf合集可在机器学习与推荐算法公众号后台回复"RecSys2020"获取。

更多干货, 欢迎大家关注仙女知乎：萌酱还是萌萌子。