[论文解读] The closed loop between opinion formation and personalised recommendations

Wilbert Samuel Rossi, Jan Willem Polderman+1 more|arXiv (Cornell University)|Sep 12, 2018

Opinion Dynamics and Social Influence参考文献 36被引用 8

一句话总结

本文提出了一种可计算的数学模型，用于描述用户意见形成与个性化新闻推荐之间的反馈回路，表明推荐系统通过放大确认偏误，会促使用户观点趋于极端。关键发现是一种可量化的权衡关系：个性化推荐带来的点击率提升，是以意见扭曲加剧为代价的，该关系通过意见偏离与参与度增益之间的解析关系得以形式化。

ABSTRACT

In online platforms, recommender systems are responsible for directing users to relevant contents. In order to enhance the users' engagement, recommender systems adapt their output to the reactions of the users, who are in turn affected by the recommended contents. In this work, we study a tractable analytical model of a user that interacts with an online news aggregator, with the purpose of making explicit the feedback loop between the evolution of the user's opinion and the personalised recommendation of contents. More specifically, we assume that the user is endowed with a scalar opinion about a certain issue and seeks news about it on a news aggregator: this opinion is influenced by all received news, which are characterized by a binary position on the issue at hand. The user is affected by a confirmation bias, that is, a preference for news that confirm her current opinion. The news aggregator recommends items with the goal of maximizing the number of user's clicks (as a measure of her engagement): in order to fulfil its goal, the recommender has to compromise between exploring the user's preferences and exploiting what it has learned so far. After defining suitable metrics for the effectiveness of the recommender systems (such as the click-through rate) and for its impact on the opinion, we perform both extensive numerical simulations and a mathematical analysis of the model. We find that personalised recommendations markedly affect the evolution of opinions and favor the emergence of more extreme ones: the intensity of these effects is inherently related to the effectiveness of the recommender. We also show that by tuning the amount of randomness in the recommendation algorithm, one can seek a balance between the effectiveness of the recommendation system and its impact on the opinions.

研究动机与目标

• 建模在线新闻平台中用户意见演化与个性化推荐系统之间的闭环互动。
• 量化确认偏误与推荐策略如何共同影响意见极端化与用户参与度。
• 分析推荐有效性（点击率）与意见扭曲影响之间的权衡关系。
• 探讨通过调节推荐算法中的随机性，能否在参与度与意见稳定性之间取得平衡。

提出的方法

• 一种标量意见模型，用户根据接收到的具有二元立场（+1 或 -1）的新闻条目更新其观点，立场针对某一政治或社会议题。
• 用户表现出确认偏误，当新闻立场与其当前观点一致时，点击概率随之增加。
• 推荐系统采用 epsilon-greedy 策略，在探索（测试未知立场）与利用（推荐高点击项目）之间取得平衡，以最大化点击次数。
• 用户的意见演化为一个仿射动力系统，随时间整合新闻信息，其参数控制学习速率与噪声水平。
• 推导出稳态意见与点击率的解析解，从而实现对意见扭曲与参与度增益的量化分析。
• 数值模拟验证了在不同推荐偏置水平（epsilon）与用户初始观点下的解析结果。

实验结果

研究问题

• RQ1 在存在确认偏误的情况下，个性化推荐如何影响用户长期意见的演化？
• RQ2 推荐系统性能（点击率）与其对意见极端化的影响之间存在何种定量关系？
• RQ3 通过调节推荐算法中的探索-利用权衡，能在多大程度上缓解意见扭曲？
• RQ4 初始用户意见与推荐偏置参数 epsilon 如何影响最终意见与参与度？

主要发现

• 个性化推荐显著加剧了意见极端化，意见扭曲 Δusr± 随 (1−2ε) 线性增加，其中 ε 控制推荐偏置。
• 点击率增益 Γctr± 随意见扭曲增加而上升，其解析关系为 Γctr± = ½(α/γ)o_usr⁰Δusr± + ½((α+γ)/γ)(Δusr±)²，表明参与度与意见偏离之间存在直接权衡。
• 当 ε ≠ 0.5 时，系统会产生持续的意见扭曲 Δusr± = ±(γ/(α+γ))(1−2ε)，意味着即使推荐中存在微小偏置，也会导致可测量的意见偏移。
• 该模型证实，更高的推荐有效性（通过利用策略实现）会导致更严重的意见极化，而增加探索则可减少扭曲，代价是参与度降低。
• 模拟结果验证了分析预测，显示在不同参数设置下，实证点击率增益与预测意见扭曲之间具有高度一致性。