一文纵览多样化推荐的现状与发展
© 作者|徐澜玲
机构|中国人民大学
引言:推荐系统旨在挖掘用户兴趣并为用户提供个性化的推荐服务。然而,只关注准确性的推荐算法对用户、供应商和推荐平台都有一定的负面影响。为此,公平性、隐私性、多样性等准确性外的指标也引起了研究者们的关注。本文聚焦于推荐系统中的个体多样性,将推荐系统个体多样性的相关工作分为三类进行介绍:前处理方法、中处理方法和后处理方法。文章也同步发布在 AI Box 知乎专栏(知乎搜索 AI Box 专栏),欢迎大家在知乎专栏的文章下方评论留言,交流探讨!
图 1: 多样化推荐系统的方法分类:前处理、中处理和后处理方法
https://github.com/RUCAIBox/RecBole
1. 研究背景
1.1 推荐系统为什么需要多样性
Customer: 对于客户来说,用户本身可能有多方面的兴趣,但对不同领域/类型的偏好存在差异。然而,推荐算法端训练时只强调了用户的部分主流兴趣,导致头部过拟合、尾部欠拟合的问题。久而久之,过曝光效应导致了过滤气泡和信息茧房的问题,用户逐渐对推荐结果感到乏味,影响客户留存率和满意度(图 2)。 Provider: 对于商家而言,他们可能会面临曝光分布不均的问题,即优秀或受欢迎的供应商将占据市场的主导地位,威胁小商家或新供应商的生存,垄断的商家格局将影响推荐平台的长期发展。
1.2 个体 v.s. 系统级别的多样性
个人层面的多样性主要集中在为顾客提供多样化的推荐结果。以新闻推荐为例 [41],同质化新闻会强化用户已有的认知,产生极化和扩散效应。此外,奢侈品时尚等一般性推荐的研究 [53] 也表明发现独家和稀有商品对于提高用户粘性的重要性。因此,多样化推荐结果对用户自身有积极的影响,避免单调乏味的平台体验。
系统层面的多样性指的是不同供应商之间的公平曝光 [59, 70]。在这种定义下,公平性和长尾推荐都可以表示为系统级的多样性。
1.3 多样性与相关概念的区别和联系
多样性 (diversity):推荐结果的多样化程度,多样的推荐结果期望结果间的平均相似度低。而 coverage (覆盖度) 指的是推荐结果对物品集的覆盖程度,一般可以作为多样性的衡量指标。 公平性 (fairness):推荐系统中的公平性可以按照作用的对象分为个体公平和组公平。个体公平要求相似的个体应该得到相似的推荐结果,而组公平要求不同的群体应该得到平等的对待。当我们考虑组公平和系统级别的多样性时,公平性和多样性都涉及到对弱势组的考量,具有相似的目的。例如,在推荐的场景下,供应商的公平性 (provider fairness) 可以视作系统级的多样性。 惊喜性 (serendipity):惊喜性有时也称作意外度 (unexpectedness),表示用户对于推荐物品的惊喜程度,可以用推荐物品与历史交互之间的距离来衡量。若算法给出的物品是用户可能感兴趣但之前未曾涉猎的范围,则惊喜性具有帮助用户跳出信息茧房的潜力。 新颖性 (novelty):新颖性指的是推荐结果中有新 (novel) 的物品,但对于新颖性的衡量和定义并没有统一的标准。新颖性与惊喜性有一定的相似之处,两者对提升多样性都有一定的帮助。相比之下,多样性是更通用、含义更丰富的概念。
2. 多样性的评测指标
2.1 覆盖度
2.2 列表内距离
2.3 熵
2.4 基尼系数
2.5 α−NDCG
3. 前处理方法
类别 | 方法 |
3.1 预先定义用户类型
3.2 基于相似用户的数据增广
图 4: A Relevant and Diverse Retrieval-enhanced Data Augmentation Framework for Sequential Recommendation [6]
3.3 基于图神经网络的数据增广
图 5: Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering [61]
3.4 前处理方法小结
4. 中处理方法
类别 | 方法 |
4.1 基于正则化的方法
4.2 基于模型框架的方法
4.3 基于强化学习的方法
4.4 基于因果的方法
图 6: [64] 分析了过滤气泡和推荐算法的偏差放大效应
4.5 中处理方法小结
5. 后处理方法
类别 | 方法 |
5.1 基于最大边际相关性的方法
图 7: Feature-aware Diversified Re-ranking with Disentangled Representations for Relevant Recommendation [40]
5.2 基于行列式点过程的方法
5.3 基于精炼的方法
5.4 后处理方法小结
6. 总结与未来展望
6.1 不同方法的全面评测
6.2 多兴趣建模中的多样性
6.3 可微的多样性评测指标
6.4 准确性和多样性的权衡
6.5 可解释的多样化推荐方法
6.6 考虑多方的多样化推荐
参考文献
[1] M. Abdool et al. Managing Diversity in Airbnb Search. SIGKDD 2020. [2] G. Adomavicius et al. Maximizing Aggregate Recommendation Diversity: A graph Theoretic Approach. DiveRS 2011. [3] A. Antikacioglu et al. Post Processing Recommender Systems for Diversity. SIGKDD 2017. [4] A. Ashkan et al. Optimal Greedy Diversity for Recommendation. IJCAI 2015. [5] G. Balloccu et al. Post Processing Recommender Systems with Knowledge Graphs for Recency, Popularity, and Diversity of Explanations. SIGIR 2022. [6] S. Bian et al. A Relevant and Diverse Retrieval-enhanced Data Augmentation Framework for Sequential Recommendation. CIKM 2022. [7] M. Boeker et al. An Empirical Investigation of Personalization Factors on TikTok. WWW 2022. [8] R. Boim et al. Diversification and Refinement in Collaborative Filtering Recommender. CIKM 2011 [9] J. Carbonell et al. The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries. SIGIR 1998. [10] Y. Cen et al. Controllable Multi-Interest Framework for Recommendation. SIGKDD 2020. [11] O. Chapelle et al. Expected Reciprocal Rank for Graded Relevance. CIKM 2009. [12] L. Chen et al. Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity. NeurIPS 2018. [13] W. Chen et al. Improving End-to-End Sequential Recommendations with Intent-aware Diversification. CIKM 2020. [14] P. Cheng et al. Learning to Recommend Accurate and Diverse Items. WWW 2017. [15] C. L. Clarke et al. Novelty and Diversity in Information Retrieval Evaluation. SIGIR 2008. [16] C. L. Clarke et al. An Effectiveness Measure for Ambiguous and Underspecified Queries. ICTIR 2009. [17] T. Di Noia et al. An Analysis of Users' Propensity Toward Diversity in Recommendations. RecSys 2014. [18] Q. Ding et al. A Hybrid Bandit Framework for Diversified Recommendation. AAAI 2021. [19] F. Eskandanian et al. A Clustering Approach for Personalizing Diversity in Collaborative Recommender Systems. UMAP 2017. [20] D. M. Fleder et al. Recommender Systems and Their Impact on Sales Diversity. EC 2007. [21] L. Gan et al. Enhancing Recommendation Diversity using Determinantal Point Processes on Knowledge Graphs. SIGIR 2020. [22] Z. Gao et al. Mitigating the Filter Bubble While Maintaining Relevance: Targeted Diversification with VAE-based Recommender Systems. SIGIR 2022. [23] Y. Ge et al. Understanding Echo Chambers in E-commerce Recommender Systems. SIGIR 2020. [24] S. Gong et al. Positive, Negative and Neutral: Modeling Implicit Feedback in Session-based News Recommendation. SIGIR 2022. [25] W. Gong et al. DAWAR: Diversity-aware Web APIs Recommendation for Mashup Creation based on Correlation Graph. SIGIR 2022. [26] J. Hao et al. P-Companion: A Principled Framework for Diversified Complementary Product Recommendation. CIKM 2020. [27] K. Hirata et al. Solving Diversity-Aware Maximum Inner Product Search Efficiently and Effectively. RecSys 2022. [28] Y. Huang et al. Sliding Spectrum Decomposition for Diversified Recommendation. SIGKDD 2021. [29] N. J. Hurley. Personalised Ranking with Diversity. RecSys 2013. [30] K. Järvelin et al. Cumulated Gain-based Evaluation of IR Techniques. TOIS 2002. [31] Y. Kim et al. Sequential and Diverse Recommendation with Long Tail. IJCAI 2019. [32] H. Kwon et al. ART (Attractive Recommendation Tailor): How the Diversity of Product Recommendations Affects Customer Purchase Preference in Fashion Industry? CIKM 2020. [33] N. Lathia et al. Temporal Diversity in Recommender Systems. SIGIR 2010. [34] R. I. Lerman et al. A Note on the Calculation and Interpretation of the Gini Index. Economics Letters 1984. [35] C. Li et al. Cascading Hybrid Bandits: Online Learning to Rank for Relevance and Diversity. RecSys 2020. [36] J. Li et al. FairGAN: GANs-based Fairness-aware Learning for Recommendations with Implicit Feedback. WWW 2022. [37] N. Li et al. An Exploratory Study of Information Cocoon on Short-form Video Platform. CIKM 2022. [38] P. Li et al. PURS: Personalized Unexpected Recommender System for Improving User Satisfaction. RecSys 2020. [39] Y. Liang et al. Enhancing Domain-Level and User-Level Adaptivity in Diversified Recommendation. SIGIR 2021. [40] Z. Lin et al. Feature-aware Diversified Re-ranking with Disentangled Representations for Relevant Recommendation. SIGKDD 2022. [41] P. Liu et al. The Interaction between Political Typology and Filter Bubbles in News Recommendation Algorithms. WWW 2021. [42] S. Liu et al. Long-tail Session-based Recommendation. RecSys 2020. [43] Y. Liu et al. Determinantal Point Process Likelihoods for Sequential Recommendation. SIGIR 2022. [44] Y. Liu et al. Diversified Interactive Recommendation with Implicit Feedback. AAAI 2020. [45] Y. Lu et al. Future-Aware Diverse Trends Framework for Recommendation. WWW 2021. [46] J. Parapar et al. Diverse User Preference Elicitation with Multi-Armed Bandits. WSDM 2021. [47] J. Parapar et al. Towards Unified Metrics for Accuracy and Diversity for Recommender Systems. RecSys 2021. [48] B. Paudel et al. Random Walks with Erasure: Diversifying Personalized Recommendations on Social and Information Networks. WWW 2021. [49] C. Pei et al. Personalized Re-ranking for Recommendation. RecSys 2019. [50] L. Peska et al. Towards Results-level Proportionality for Multi-objective Recommender Systems. SIGIR 2022. [51] S. A. Puthiya Parambath. A Coverage-Based Approach to Recommendation Diversity On Similarity Graph. RecSys 2016. [52] L. Qin et al. Promoting Diversity in Recommendation by Entropy Regularizer. IJCAI 2013. [53] J. Sá et al. Diversity Vs Relevance: A Practical Multi-objective Study in Luxury Fashion Recommendations. SIGIR 2022. [54] C. Sha et al. A Framework for Recommending Relevant and Diverse Items. IJCAI 2016. [55] Y. Shi et al. Adaptive Diversification of Recommendation Results via Latent Factor Portfolio. SIGIR 2012. [56] B. Smyth et al. Similarity vs. Diversity. LNAI 2001. [57] I. Srba et al. Auditing YouTube’s Recommendation Algorithm for Misinformation Filter Bubbles. TOIS 2023. [58] D. Stamenkovic et al. Choosing the Best of Both Worlds: Diverse and Novel Recommendations through Multi-Objective Reinforcement Learning. WSDM 2022. [59] H. Steck. Item Popularity and Recommendation Accuracy. RecSys 2011. [60] J. Sun et al. A Framework for Recommending Accurate and Diverse Items Using Bayesian Graph Convolutional Neural Networks. SIGKDD 2020. [61] C. Tian et al. Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering. SIGIR 2022. [62] S. Vargas et al. Coverage, Redundancy and Size-awareness in Genre Diversity for Recommender Systems. RecSys 2014. [63] S. Vargas et al. Intent-oriented Diversity in Recommender Systems. SIGIR 2011. [64] W. Wang et al. Deconfounded Recommendation for Alleviating Bias Amplification. SIGKDD 2021. [65] W. Wang et al. User-controllable Recommendation Against Filter Bubbles. SIGIR 2022. [66] M. Wilhelm et al. Practical Diversified Recommendations on YouTube with Determinantal Point Processes. CIKM 2018. [67] C. Wu et al. Is News Recommendation a Sequential Recommendation Task? SIGIR 2022. [68] H. Wu et al. A Survey of Diversification Techniques in Search and Recommendation. arxiv 2022. [69] Q. Wu et al. PD-GAN: Adversarial Learning for Personalized Diversity-Promoting Recommendation. IJCAI 2019. [70] Y. Wu et al. TFROM: A Two-sided Fairness-Aware Recommendation Model for Both Customers and Providers. SIGIR 2021. [71] W. Yang et al. On the Diversity and Explainability of Recommender Systems: A Practical Framework for Enterprise App Recommendation. CIKM 2021. [72] R. Ye et al. Dynamic Graph Construction for Improving Diversity of Recommendation. RecSys 2021. [73] M. Zhang et al. Avoiding Monotony: Improving the Diversity of Recommendation Lists. RecSys 2008. [74] X. Zhang et al. Disentangled Representation for Diversified Recommendations. arxiv 2023. [75] M. Zhao et al. Investigating Accuracy-Novelty Performance for Graph-based Collaborative Filtering. SIGIR 2022. [76] Y. Zheng et al. DGCN: Diversified Recommendation with Graph Convolutional Networks. WWW 2021. [77] C.-N. Ziegler et al. Improving Recommendation Lists through Topic Diversification. WWW 2005. [78] S. Yang et al. Graph Exploration Matters: Improving both individual-level and system-level diversity in WeChat Feed Recommender. arxiv 2023.
更多推荐
↓↓↓