Recsys2021 | 推薦系統論文整理和導讀

本期主要整理和分類了Recsys 2021的Research Papers和Reproducibility papers。按照推薦系統的研究方向和使用的推薦技術來分類，友善大家快速檢索自己感興趣的文章。個人認為Recsys這個會議重點不在于"技術味多濃"或者"技術多先進"，而在于經常會湧現很多新的觀點以及有意思的研究點，涵蓋推薦系統的各個方面，例如，Recsys 2021涵蓋的一些很有意思的研究點包括：

• 推薦系統的資訊繭房和回音室問題的探讨，有4篇文章探讨了社交媒體推薦、音樂推薦和視訊推薦中的資訊繭房和回音室效應。很少見到在學術會議上專門讨論這樣深刻的問題，值得一讀。
• 推薦系統評估體系的探讨，對推薦系統整個評估體系的梳理，多個名額間如何做權衡等。
• 推薦系統的互動設計探讨，探讨了美食推薦場景下使用者互動設計。關于使用者界面/互動設計的推薦系統文章還是很新奇的。
• 推薦系統中的探索與利用探讨，例如Google關于使用者探索的工作Values of User Exploration in Recommender Systems值得一讀。
• 對已有工作的探讨和挑戰，傳統矩陣分解推薦系統和深度學習推薦系統的對比。例如：何向南老師的NCF工作和MF的對比，繼Recsys20被進行對比後, 在Recsys21上又再次被擺上了台面進行對比。
  • Recsys20, Rendle S, Krichene W, Zhang L, et al. Neural collaborative filtering vs. matrix factorization revisited[C]//Fourteenth ACM Conference on Recommender Systems. 2020: 240-248.
  • Recsys21, Anelli V W, Bellogín A, Di Noia T, et al. Reenvisioning the comparison between Neural Collaborative Filtering and Matrix Factorization[C]//Fifteenth ACM Conference on Recommender Systems. 2021: 521-529.

還有些研究點也是值得一讀的，比如推薦系統中的冷啟動，偏差與糾偏，序列推薦，可解釋性，隐私保護等，這些研究很有意思和啟發性，有助于開拓大家的研究思路。

下面主要根據自己讀題目或者摘要時的一些判斷做的歸類，按照推薦系統研究方向分類、推薦技術分類以及專門實驗性質的可複現型文章分類，可能存在漏歸和錯歸的情況，請大家多多指正。

1.按照推薦系統研究方向分類

1.1 資訊繭房和回音室

資訊繭房/回音室(echo chamber)/過濾氣泡(filter bubble)，這3個概念類似，在國内外有不同的說法。大緻是指使用社交媒體以及帶有算法推薦功能的資訊類APP，可能會導緻我們隻看得到自己感興趣的、認同的内容，進而讓大家都活在自己的小世界裡，彼此之間難以認同和溝通。關于這部分的概念可參見知乎文章：https://zhuanlan.zhihu.com/p/71844281。有四篇文章探讨了這樣的問題。

• The Dual Echo Chamber: Modeling Social Media Polarization for Interventional Recommending

  Tim Donkers and Jürgen Ziegler

• I want to break free! Recommending friends from outside the echo chamber

  Antonela Tommasel, Juan Manuel Rodriguez, and Daniela Godoy

• Follow the guides: disentangling human and algorithmic curation in online music consumption

  Quentin Villermet, Jérémie Poiroux, Manuel Moussallam, Thomas Louail, and Camille Roth

• An Audit of Misinformation Filter Bubbles on YouTube: Bubble Bursting and Recent Behavior Changes

  Matus Tomlein, Branislav Pecher, Jakub Simko, Ivan Srba, Robert Moro, Elena Stefancova, Michal Kompan, Andrea Hrckova, Juraj Podrouzek, and Maria Bielikova

1.2 探索與利用

此次大會在探索與利用上也有很多探讨，例如多臂老虎機、谷歌的新工作，即：使用者側的探索等。

• Burst-induced Multi-Armed Bandit for Learning Recommendation

  Rodrigo Alves, Antoine Ledent, and Marius Kloft

• Values of User Exploration in Recommender Systems

  Google, Minmin Chen, Yuyan Wang, Can Xu, Ya Le, mohit sharma, Lee Richardson, and Ed Chi

• Designing Online Advertisements via Bandit and Reinforcement Learning

  Yusuke Narita, Shota Yasui, and Kohei Yata

• The role of preference consistency, defaults and musical expertise in users’ exploration behavior in a genre exploration recommender

  Yu Liang and Martijn C. Willemsen

• Top-K Contextual Bandits with Equity of Exposure

  Olivier Jeunen and Bart Goethals

1.3 偏差與糾偏

涉及排序學習的糾偏、使用者的偏差探索等。

Debiased Explainable Pairwise Ranking from Implicit Feedback

Khalil Damak, Sami Khenissi, and Olfa Nasraoui

Mitigating Confounding Bias in Recommendation via Information Bottleneck

Dugang Liu, Pengxiang Cheng, Hong Zhu, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming

User Bias in Beyond-Accuracy Measurement of Recommendation Algorithms

Ningxia Wang, and Li Chen

1.4 冷啟動

利用圖學習、表征學習等做冷啟動。

Cold Start Similar Artists Ranking with Gravity-Inspired Graph Autoencoders

Guillaume Salha-Galvan, Romain Hennequin, Benjamin Chapus, Viet-Anh Tran, and Michalis Vazirgiannis

Shared Neural Item Representations for Completely Cold Start Problem

Ramin Raziperchikolaei, Guannan Liang, and Young-joo Chung

1.5 評估體系

涉及離線或線上評估方法，準确性和多樣性等統一名額的設計等。

Evaluating Off-Policy Evaluation: Sensitivity and Robustness

Yuta Saito, Takuma Udagawa, Haruka Kiyohara, Kazuki Mogi, Yusuke Narita, and Kei Tateno

Fast Multi-Step Critiquing for VAE-based Recommender Systems

Diego Antognini and Boi Faltings

Online Evaluation Methods for the Causal Effect of Recommendations

Masahiro Sato

Towards Unified Metrics for Accuracy and Diversity for Recommender Systems

Javier Parapar and Filip Radlinski

1.6 會話/序列推薦

涉及session次元的短序列推薦；使用NLP中常用的Transformers做序列推薦的鴻溝探讨和解決，這個工作本人還挺感興趣的，後續會精讀下！

• Next-item Recommendations in Short Sessions

  Wenzhuo Song, Shoujin Wang, Yan Wang, and SHENGSHENG WANG

• Transformers4Rec: Bridging the Gap between NLP and Sequential / Session-Based Recommendation

  Gabriel de Souza Pereira Moreira, Sara Rabhi, Jeong Min Lee, Ronay Ak, and Even Oldridge

• Denoising User-aware Memory Network for Recommendation

  Zhi Bian, Shaojun Zhou, Hao Fu, Qihong Yang, Zhenqi Sun, Junjie Tang, Guiquan Liu, kaikui liu, and Xiaolong Li

• Large-Scale Modeling of Mobile User Click Behaviors Using Deep Learning

  Xin Zhou and Yang Li

1.7 隐私保護

結合聯邦學習做隐私保護等。

• Privacy Preserving Collaborative Filtering by Distributed Mediation

  Alon Ben Horin, and Tamir Tassa

• Stronger Privacy for Federated Collaborative Filtering With Implicit Feedback

  Lorenzo Minto, Moritz Haller, Ben Livshits, and Hamed Haddadi

1.8 對抗與攻擊

Black-Box Attacks on Sequential Recommenders via Data-Free Model Extraction

Zhenrui Yue, Zhankui He, Huimin Zeng, and Julian McAuley

1.9 對話推薦系統

Large-scale Interactive Conversational Recommendation System

Ali Montazeralghaem, James Allan, and Philip S. Thomas

1.10 可解釋性推薦

EX3: Explainable Attribute-aware Item-set Recommendations

Yikun Xian, Tong Zhao, Jin Li, Jim Chan, Andrey Kan, Jun Ma, Xin Luna Dong, Christos Faloutsos, George Karypis, S. Muthukrishnan, and Yongfeng Zhang

1.11 跨域推薦

Towards Source-Aligned Variational Models for Cross-Domain Recommendation

Aghiles Salah, Thanh Binh Tran, and Hady Lauw

1.12 基于視覺的推薦

利用視覺資訊做推薦。

• Semi-Supervised Visual Representation Learning for Fashion Compatibility

Ambareesh Revanur, Vijay Kumar, and Deepthi Sharma

• Tops, Bottoms, and Shoes: Building Capsule Wardrobes via Cross-Attention Tensor Network

Huiyuan Chen, Yusan Lin, Fei Wang, and Hao Yang

1.13 組推薦/使用者物品分層推薦

• Local Factor Models for Large-Scale Inductive Recommendation

  Longqi Yang, Tobias Schnabel, Paul N. Bennett, and Susan Dumais

• Learning to Represent Human Motives for Goal-directed Web Browsing

  Jyun-Yu Jiang, Chia-Jung Lee, Longqi Yang, Bahareh Sarrafzadeh, Brent Hecht, Jaime Teevan

1.14 推薦系統互動設計

探讨了美食場景下，多使用者意圖的推薦系統的互動設計。

“Serving Each User”: Supporting Different Eating Goals Through a Multi-List Recommender Interface

Alain Starke, Edis Asotic, and Christoph Trattner

2. 按照推薦技術分類

涉及傳統協同過濾、度量學習的疊代；新興的圖學習技術、聯邦學習技術、強化學習技術等的探索。

2.1 協同過濾

探索了傳統的協同過濾工作，其中第一篇工作把CF和LDA聯系在了一起，挺有意思。

Matrix Factorization for Collaborative Filtering Is Just Solving an Adjoint Latent Dirichlet Allocation Model After All

Florian Wilhelm

Negative Interactions for Improved Collaborative-Filtering: Don’t go Deeper, go Higher

Harald Steck and Dawen Liang

ProtoCF: Prototypical Collaborative Filtering for Few-shot Item Recommendation

Aravind Sankar, Junting Wang, Adit Krishnan, and Hari Sundaram

2.2 圖學習

知識圖譜的應用以及圖嵌入技術和上下文感覺的表征技術的融合，這兩個工作個人都挺感興趣。

• Sparse Feature Factorization for Recommender Systems with Knowledge Graphs

 Antonio Ferrara, Vito Walter Anelli, Tommaso Di Noia, and Alberto Carlo            Maria Mancino

• Together is Better: Hybrid Recommendations Combining Graph Embeddings and Contextualized Word Representations

 Marco Polignano, Cataldo Musto, Marco de Gemmis, Pasquale Lops, and       Giovanni Semeraro

2.3 強化學習

強化學習在推薦系統中的應用，和對話系統結合在一起；獎勵函數的設計等。

• Partially Observable Reinforcement Learning for Dialog-based Interactive Recommendation

  Yaxiong Wu, Craig Macdonald, and Iadh Ounis,

• Pessimistic Reward Models for Off-Policy Learning in Recommendation

Olivier Jeunen and Bart Goethals

2.4 度量學習

協同過濾和度量學習的結合，即：CML。

• Hierarchical Latent Relation Modeling for Collaborative Metric Learning

  Viet-Anh Tran, Guillaume Salha-Galvan, Romain Hennequin, and Manuel Moussallam

2.5 聯邦學習

聯邦學習的優化以及在隐私保護中的應用。

• A Payload Optimization Method for Federated Recommender Systems

  Farwa K. Khan, Adrian Flanagan, Kuan Eeik Tan, Zareen Alamgir, and Muhammad Ammad-ud-din

• Stronger Privacy for Federated Collaborative Filtering With Implicit Feedback

     Lorenzo Minto, Moritz Haller, Ben Livshits, and Hamed Haddadi

2.6 架構/訓練/優化

涉及訓練、優化、檢索、實時流等。

• cDLRM: Look Ahead Caching for Scalable Training of Recommendation Models

  Keshav Balasubramanian, Abdulla Alshabanah, Joshua D Choe, and Murali Annavaram

• Reverse Maximum Inner Product Search: How to efficiently find users who would like to buy my item?

  Daichi Amagata and Takahiro Hara

• Page-level Optimization of e-Commerce Item RecommendationsChieh Lo, Hongliang Yu, Xin Yin, Krutika Shetty, Changchen He, Kathy Hu, Justin M Platz, Adam Ilardi, and Sriganesh Madhvanath

• Accordion: A Trainable Simulator for Long-Term Interactive Systems

  James McInerney, Ehtsham Elahi, Justin Basilico, Yves Raimond, and Tony Jebara

• Information Interactions in Outcome Prediction: Quantification and Interpretation using Stochastic Block Models

  Gaël Poux-Médard, Julien Velcin, and Sabine Loudcher

• Learning An Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems

  Danni Peng, Sinno Jialin Pan, Jie Zhang, and Anxiang Zeng

• Recommendation on Live-Streaming Platforms: Dynamic Availability and Repeat Consumption

Jeremie Rappaz, Julian McAuley, and Karl Aberer

3. 實驗性質的文章

Reproducibility papers可複現實驗性質的文章，共3篇。分别探索了：序列推薦中的采樣評估政策；對話推薦系統中生成式和檢索式的方法對比；神經網絡推薦系統和矩陣分解推薦系統的對比。

• A Case Study on Sampling Strategies for Evaluating Neural Sequential Item Recommendation Models

  by Alexander Dallmann, Daniel Zoller, Andreas Hotho (Data Science Chair, University of Würzburg, Würzburg, Germany)

• Generation-based vs. Retrieval-based Conversational Recommendation: A User-Centric Comparison

  by Ahtsham Manzoor and Dietmar Jannach (University of Klagenfurt, Klagenfurt, Austria)

• Reenvisioning the comparison between Neural Collaborative Filtering and Matrix Factorization

  by Vito Walter Anelli (Polytechnic University of Bari, Bari, Italy), Alejandro Bellogin (Information Retrieval Group, Universidad Autonoma de Madrid, Madrid, Spain), Tommaso Di Noia Polytechnic (University of Bari, Bari, Italy), and Claudio Pomo (Polytechnic University of Bari, Bari, Italy)

總結

通過此次的論文的整理和分類，筆者也發現了一些自己感興趣的研究點，比如：推薦系統的回音室效應探讨文章；Transformers在序列推薦和NLP序清單征中的鴻溝和解決文章：Transformers4Rec；圖嵌入表征和上下文感覺表征的融合文章；NCF和MF的實驗對比文章；谷歌的使用者探索文章等。希望讀者也能夠發現自己感興趣的文章。下期分享見！