Filter Bubbles in Recommender Systems: Fact or Fallacy - A Systematic Review
推荐系统中的过滤气泡：事实还是谬误 - 系统回顾

Recommender systems (RSs) play a crucial role in providing personalized suggestions to users based on their past interactions. These systems encompass a wide range of recommendations, including movies, products, travel options, advertisements, and news. User preferences can be inferred from their behavior, which can be either implicit or explicit. Implicit preferences are deduced from activities such as online shopping, website visits, link clicks, and web browser cookies, without directly soliciting feedback from users. On the other hand, explicit feedback involves actively requesting users to provide ratings or comments on the recommendations they have received [25]. Content-based filtering, collaborative filtering, and hybrid approaches are the three most commonly employed recommendation techniques in RSs [26, 27]. Commercial recommendation methods often adopt a combination of these approaches rather than relying solely on content or collaborative filtering. They frequently integrate knowledge-based and context-based strategies to enhance the accuracy and effectiveness of recommendations [26].
推荐系统（RS）在根据用户过去的交互向用户提供个性化建议方面发挥着至关重要的作用。这些系统涵盖广泛的推荐，包括电影、产品、旅行选项、广告和新闻。用户偏好可以从他们的行为中推断出来，可以是隐式的，也可以是显式的。隐式偏好是从在线购物、网站访问、链接点击和网络浏览器 cookie 等活动中推断出来的，而不直接征求用户的反馈。另一方面，显式反馈涉及主动请求用户对他们收到的推荐提供评级或评论[25]。基于内容的过滤、协同过滤和混合方法是推荐系统中最常用的三种推荐技术[26, 27]。商业推荐方法通常采用这些方法的组合，而不是仅仅依赖内容或协同过滤。他们经常整合基于知识和基于情境的策略来提高建议的准确性和有效性[26]。

The distinction between a current experience and one that has already occurred can be described as novelty, while the internal variations within the components of an experience are referred to as diversity. Initially, recommender systems (RSs) were primarily designed to predict users’ interests. However, as research on RSs progressed, the literature began to emphasize a broader perspective on recommendation utility, which includes not only prediction accuracy [28, 29], but also the importance of originality, variety, and other features in enhancing the value of recommendations [30, 31]. This awareness has grown over time, leading to a surge of activity in this area over the past decade [32, 33, 34, 35, 36]. As a result, novelty and diversity have gained prominence and are increasingly recognized as important evaluation measures for new recommendation systems. Algorithmic advancements are consistently aimed at improving these aspects.
当前体验和已经发生的体验之间的区别可以被描述为新颖性，而体验组成部分内的内部变化被称为多样性。最初，推荐系统（RS）主要是为了预测用户的兴趣而设计的。然而，随着RS研究的进展，文献开始强调推荐效用的更广泛视角，其中不仅包括预测准确性[28, 29]，还包括原创性、多样性和其他特征在增强推荐价值方面的重要性[ 30, 31]。随着时间的推移，这种意识不断增强，导致过去十年该领域的活动激增[32,33,34,35,36]。因此，新颖性和多样性受到重视，并越来越被认为是新推荐系统的重要评估指标。算法的进步始终致力于改善这些方面。

In recent decades, the rise of the Internet has sparked considerable scholarly interest in its potential negative effects on society and the public sphere [37]. The concept of the internet filter bubble has gained widespread recognition as a manifestation of this pessimistic perspective. The underlying premise of an echo chamber is that social media users deliberately interact with like-minded individuals and consume content that aligns with their ideologies. As a result, they rarely encounter diverse viewpoints that are crucial for fostering a more inclusive and vibrant public sphere [38].
近几十年来，互联网的兴起引发了学术界对其对社会和公共领域潜在负面影响的极大兴趣[37]。作为这种悲观观点的体现，互联网过滤泡沫的概念已获得广泛认可。回声室的基本前提是社交媒体用户故意与志同道合的人互动，并消费符合他们意识形态的内容。因此，他们很少遇到对于培育更具包容性和活力的公共领域至关重要的不同观点[38]。

This phenomenon is exacerbated by the algorithmic content selection employed by social media platforms, which tends to limit users’ exposure to novel and diverse content. As a result, online communities become clustered and polarized, lacking the necessary viewpoint diversity. The concept of the "Filter Bubble" refers to the potential consequence of personalized internet customization, where individuals are isolated from diverse perspectives and information. Users often find themselves exposed to familiar content or consistent information on similar topics, reinforcing their existing knowledge. This concern initially arose in 2009 when platforms like Google began prioritizing customized search results, leading to variations in outcomes for users based on their previous interactions, expressed preferences, and other criteria [26].
社交媒体平台采用的算法内容选择加剧了这种现象，这往往会限制用户接触新颖且多样化的内容。结果，在线社区变得集群化和两极分化，缺乏必要的观点多样性。 “过滤泡沫”的概念指的是个性化互联网定制的潜在后果，即个人与不同的观点和信息隔离。用户经常发现自己接触到熟悉的内容或类似主题的一致信息，从而强化了他们现有的知识。这种担忧最初出现在 2009 年，当时 Google 等平台开始优先考虑定制搜索结果，导致用户的结果根据他们之前的互动、表达的偏好和其他标准而变化 [26]。

Consumers now encounter a more personalized online environment that delivers content tailored to their perceived interests and the preferences of like-minded individuals within their network. While recommendation engines effectively identify users’ preferred choices, they can also contribute to information polarization and restrict novelty and variety, exerting a significant influence on user preferences and satisfaction. Consequently, users are exposed to a narrower range of information and content, as recommended and selected options are reinforced, ultimately leading to the formation of information cocoons. In the field of media communication, this phenomenon is commonly referred to as "echo chambers" [47], while information retrieval scholars label it as "filter bubbles." Filter bubbles represent self-reinforcing systems that isolate individuals from diverse ideas, beliefs, or content [48]. The filter bubble effect facilitates the solidification of existing beliefs and preferences, potentially leading to the adoption of more extreme views or behaviors over time, a phenomenon known as "group polarization" [49]. In the business context, the filter bubble effect gives rise to the "Matthew effect" among popular items, wherein products and information that deviate from the long tail hypothesis are not recommended, resulting in reduced sales diversity and potential limitations to corporate success [50, 51]. Furthermore, the prevalence of the filter bubble effect in society can lead to the polarization of political ideas and undermine democratic fairness [52, 53]. Additionally, filter bubbles indirectly contribute to the dissemination of undesirable content on online social media platforms, such as rumors and fake news [54]. Current recommendation algorithms primarily prioritize enhancing recommendation accuracy rather than promoting diverse outcomes, which is one of the factors contributing to the formation of filter bubbles [55]. While several surveys have been conducted in recent years to explore filter bubbles and recommendation algorithms, no single study comprehensively investigates all the necessary changes required in recommendation systems to address filter bubbles. Most of the research discussed in this section consists of unstructured surveys, and relevant literature pertaining to the review of filter bubbles is also included within this domain.
消费者现在遇到了一个更加个性化的在线环境，该环境可以根据他们的感知兴趣和网络中志同道合的个人的偏好提供量身定制的内容。虽然推荐引擎有效地识别用户的偏好选择，但它们也可能导致信息两极分化并限制新颖性和多样性，对用户偏好和满意度产生重大影响。因此，用户接触到的信息和内容范围更窄，推荐和选择的选项得到加强，最终导致信息茧的形成。在媒体传播领域，这种现象通常被称为“回声室”[47]，而信息检索学者将其称为“过滤气泡”。过滤气泡代表自我强化系统，将个体与不同的想法、信仰或内容隔离开来[48]。过滤气泡效应促进了现有信念和偏好的固化，随着时间的推移，可能会导致采取更极端的观点或行为，这种现象被称为“群体极化”[49]。在商业背景下，过滤泡沫效应会在热门商品中产生“马太效应”，即不推荐偏离长尾假设的产品和信息，导致销售多样性减少，对企业成功产生潜在限制[50， 51]。此外，社会中普遍存在的过滤泡沫效应会导致政治理念的两极分化，破坏民主公平[52, 53]。此外，过滤气泡间接导致在线社交媒体平台上不良内容的传播，例如谣言和虚假新闻[54]。 当前的推荐算法主要优先考虑提高推荐准确性，而不是促进多样化的结果，这是导致过滤气泡形成的因素之一[55]。尽管近年来进行了多项调查来探索过滤气泡和推荐算法，但没有一项研究全面调查推荐系统中解决过滤气泡所需的所有必要变化。本节讨论的大多数研究都是非结构化调查，与过滤气泡评论相关的相关文献也包含在该领域内。

In 2019, [39] presented a critical analysis of the "filter bubble" hypothesis, arguing that its continued emphasis has diverted scholarly attention from more pressing areas of investigation. The authors also highlight the tangible effects of the persistent use of these notions in mainstream media and political discussions, shaping societal institutions, media and communication platforms, and individual users. Traditional broadcast media’s diminishing influence in determining information exposure has given way to contemporary information filters such as recommender systems, aggregators, search engines, feed ranking algorithms, bookmarked websites, and the individuals and organizations followed on social media platforms like Twitter. Critics express concerns that the combination of these filters may isolate individuals within their own information bubbles, making it challenging to correct any false beliefs they acquire. In [40], the authors delve into the research surrounding exposure selectivity preferences and actual exposure to shed light on this topic. Furthermore, [41] presents an integrated solution model aimed at assisting users in avoiding filter bubbles within social networks. The author conducted a comprehensive literature review, identifying 571 publications from six highly regarded scientific databases. After removing irrelevant studies and conducting an in-depth analysis of the remaining publications, a recommended category of research papers was developed. This categorization serves as the basis for designing an integrated tool that incorporates previous research findings and introduces novel features to mitigate the impact of filter bubbles.
2019年，[39]对“过滤泡沫”假说进行了批判性分析，认为其持续强调已经将学术注意力从更紧迫的研究领域转移了。作者还强调了在主流媒体和政治讨论中持续使用这些概念对塑造社会机构、媒体和传播平台以及个人用户的实际影响。传统广播媒体在决定信息曝光方面的影响力逐渐减弱，取而代之的是当代信息过滤器，例如推荐系统、聚合器、搜索引擎、提要排名算法、书签网站以及在 Twitter 等社交媒体平台上关注的个人和组织。批评者担心，这些过滤器的组合可能会将个人孤立在自己的信息泡沫中，从而使纠正他们获得的任何错误信念变得困难。在[40]中，作者深入研究了围绕暴露选择性偏好和实际暴露的研究，以阐明这一主题。此外，[41]提出了一种集成解决方案模型，旨在帮助用户避免社交网络中的过滤气泡。作者进行了全面的文献综述，从 6 个备受推崇的科学数据库中找出了 571 篇出版物。在删除不相关的研究并对剩余出版物进行深入分析后，制定了推荐的研究论文类别。这种分类是设计集成工具的基础，该工具结合了之前的研究成果，并引入了新颖的功能来减轻过滤气泡的影响。

In 2021, [42] conducted a comprehensive review of scientific literature on the subject of echo chambers in social media, aiming to provide a consolidated and critical perspective on the various techniques, similarities, differences, benefits, and limitations associated with echo chambers. This review serves as a foundation for future research in this field. The authors performed a systematic review of 55 studies that examined the presence of echo chambers on social media platforms, classifying the literature and identifying common themes in the focus, techniques, and conclusions of the studies. Similarly, in their paper, [43] provide an exploratory overview of the utilization of digital echo chambers and filter bubbles in the context of nature conservation practice. They gathered data from a literature review and a digital expert poll of German conservation actors to analyze the current understanding of these phenomena. The findings indicate that these concepts are already being investigated in relation to conservation issues, particularly climate protection, and to a lesser extent, natural conservation practice. However, there is a limited understanding of the specific mechanisms underlying digital echo chambers and filter bubbles. The study highlights the urgent need for research and strategic assessment in managing and addressing these challenges in the field of nature conservation. Furthermore, [44] conducted a semi-systematic literature review to examine the digital political economy. They identified and characterized four major threats: false news, filter bubbles/echo chambers, online hate speech, and surveillance. The authors also proposed a typology of "workable solutions" to address these risks, emphasizing the tendency to adopt technological, regulatory, and culturally ingrained approaches as part of the solution.
2021年，[42]对社交媒体中回声室主题的科学文献进行了全面回顾，旨在对与回声室相关的各种技术、相似性、差异、好处和局限性提供综合和批判性的视角。该综述可作为该领域未来研究的基础。作者对 55 项研究进行了系统回顾，这些研究检查了社交媒体平台上回声室的存在，对文献进行了分类，并确定了研究焦点、技术和结论中的共同主题。同样，在他们的论文中，[43] 提供了在自然保护实践中使用数字回声室和过滤气泡的探索性概述。他们从文献综述和对德国保护参与者的数字专家调查中收集了数据，以分析当前对这些现象的理解。研究结果表明，这些概念已经在与保护问题，特别是气候保护以及较小程度上的自然保护实践相关的方面进行了调查。然而，人们对数字回声室和过滤气泡背后的具体机制了解有限。该研究强调了在管理和应对自然保护领域的这些挑战方面迫切需要进行研究和战略评估。此外，[44]进行了半系统的文献综述来研究数字政治经济学。他们识别并描述了四种主要威胁：虚假新闻、过滤气泡/回声室、在线仇恨言论和监视。 作者还提出了解决这些风险的“可行解决方案”的类型，强调了采用技术、监管和文化根深蒂固的方法作为解决方案一部分的趋势。

In [45], the authors conducted a survey of empirical research in the Netherlands to explore tailored information delivery, with a particular focus on echo chambers and filter bubbles in a global context. The study investigated the involvement of government agencies, tech businesses, and academics in addressing these issues. Currently, the Dutch journalism landscape seems to offer a diverse range of information to different citizen groups. However, the precise impact of news personalization is not fully understood, and the increasing influence of digital corporations underscores the need for further research and deeper insights. Without a comprehensive understanding of the situation, it is challenging to develop effective strategies to mitigate the potential concerns of news customization. Similarly, in [56], a qualitative approach was employed to propose new research directions on the impact of filter bubbles on democracy. The study included a comprehensive literature review and secondary data analysis. The authors argued that the emerging financial models of digital media, heavily reliant on technology companies, marketers, and the public, contribute significantly to the creation of filter bubbles. Newsrooms increasingly gather and analyze customer data for personalized information in digital advertising and subscription models. The media industry enthusiastically embraces customization, with limited critique of its negative aspects. The authors suggested that journalism has a crucial role to play in combating information bubbles by reassessing its digital economic models and raising public awareness.
在[45]中，作者对荷兰的实证研究进行了调查，以探索量身定制的信息传递，特别关注全球背景下的回声​​室和过滤气泡。该研究调查了政府机构、科技企业和学术界在解决这些问题中的参与情况。目前，荷兰的新闻业似乎为不同的公民群体提供了多样化的信息。然而，新闻个性化的确切影响尚不完全清楚，数字公司日益增长的影响力凸显了进一步研究和更深入见解的必要性。如果不全面了解情况，就很难制定有效的策略来减轻新闻定制的潜在问题。同样，在[56]中，采用定性方法提出了关于过滤气泡对民主影响的新研究方向。该研究包括全面的文献综述和二手数据分析。作者认为，新兴的数字媒体金融模式严重依赖科技公司、营销人员和公众，极大地促成了过滤泡沫的产生。新闻编辑室越来越多地收集和分析客户数据，以获取数字广告和订阅模式中的个性化信息。媒体行业热情地拥抱定制化，但对其负面影响的批评却有限。作者认为，新闻业通过重新评估其数字经济模型和提高公众意识，在打击信息泡沫方面可以发挥至关重要的作用。

The previous review of the literature reveals a significant gap in systematic and comprehensive research specifically dedicated to investigating the filter bubble phenomenon in recommendation systems. To fill this gap, we conducted an extensive and critical investigation into the presence and impact of filter bubbles in recommendation systems. This research aims to contribute to the advancement of knowledge and understanding in the fields of recommendation systems and computational social science, offering valuable insights for both researchers and practitioners. To provide a clear overview of the existing literature reviews in this area, Table I presents a summary of relevant studies, highlighting their scope, survey methodologies employed, and the number of references considered in each study. This table serves as a reference point for understanding the scope and depth of previous research efforts in this field.
之前的文献综述揭示了专门致力于研究推荐系统中过滤气泡现象的系统性和综合性研究的显着差距。为了填补这一空白，我们对推荐系统中过滤气泡的存在和影响进行了广泛而严格的调查。这项研究旨在促进推荐系统和计算社会科学领域的知识和理解的进步，为研究人员和从业者提供有价值的见解。为了清晰地概述该领域现有的文献综述，表一总结了相关研究，强调了它们的范围、采用的调查方法以及每项研究中考虑的参考文献数量。该表可作为了解该领域先前研究工作的范围和深度的参考点。

This review article aims to address several key research questions related to the filter bubble phenomenon in recommendation systems. These research questions are as follows:
本文旨在解决与推荐系统中的过滤气泡现象相关的几个关键研究问题。这些研究问题如下：

1. 1.

   Does the filter bubble exist in recommendation systems? If so, what are the reasons for its existence?

   
   1. 推荐系统中是否存在过滤气泡？如果有，它存在的理由是什么？
2. 2.

   What are the approaches used to identify the presence of a filter bubble in recommendation systems?

   
   2. 识别推荐系统中是否存在过滤气泡的方法有哪些？
3. 3.

   How can the impact of the filter bubble be mitigated or avoided in recommendation systems?

   
   3. 在推荐系统中如何减轻或避免过滤气泡的影响？

A systematic literature review (SLR), as described by Kitchenham, is a methodological approach that involves thoroughly examining and synthesizing all relevant works pertaining to a specific research topic or subject area. Systematic reviews provide an objective and comprehensive analysis of the topic by following a rigorous and transparent process that can be audited and replicated. Despite the importance of the filter bubble phenomenon in recommendation systems, a comprehensive systematic literature review specifically focusing on this topic is currently missing in the existing literature. Therefore, conducting a thorough and meticulous analysis, guided by an SLR methodology, is crucial to examine and shed light on the various assertions and findings related to the research questions stated above in an unbiased and replicable manner [46].
正如 Kitchenham 所描述的，系统文献综述 (SLR) 是一种方法论方法，涉及彻底检查和综合与特定研究主题或学科领域相关的所有相关著作。系统评审遵循严格、透明、可审计和复制的流程，对主题进行客观、全面的分析。尽管过滤气泡现象在推荐系统中很重要，但现有文献中目前缺少专门针对该主题的全面系统的文献综述。因此，在SLR方法的指导下进行彻底而细致的分析对于以公正和可复制的方式检查和阐明与上述研究问题相关的各种主张和发现至关重要[46]。

We conducted an extensive search to gather relevant literature for this systematic literature review (SLR). We focused on recent publications available in reputable scientific journals and top conference proceedings, utilizing leading academic databases. Our search covered the period from 2012 onwards, as our findings indicated that significant research on the filter bubble phenomenon emerged during this time. To ensure a comprehensive search, we employed a double-staged search strategy consisting of Phase 1 and Phase 2 (see Figure 2). In the first phase, we manually explored various search strings and their combinations using Boolean operators. Additionally, we leveraged research databases and academic search engines to access relevant literature from multiple disciplinary and publishing platforms. Figure 3 provides an overview of the databases, libraries, and search engines that were included in our search strategy.
我们进行了广泛的检索，收集相关文献以进行系统文献综述（SLR）。我们利用领先的学术数据库，重点关注著名科学期刊和顶级会议论文集上的最新出版物。我们的搜索涵盖了从 2012 年开始的时期，因为我们的研究结果表明，关于过滤气泡现象的重要研究就是在这一时期出现的。为了确保全面的搜索，我们采用了由阶段 1 和阶段 2 组成的双阶段搜索策略（见图 2）。在第一阶段，我们使用布尔运算符手动探索各种搜索字符串及其组合。此外，我们利用研究数据库和学术搜索引擎从多个学科和出版平台访问相关文献。图 3 概述了我们的搜索策略中包含的数据库、库和搜索引擎。

During the initial phase, we refined our search strings by incorporating specific keywords, including terms from the titles, abstracts, and relevant keywords of publications. This iterative process allowed us to fine-tune our search results and address any potential limitations in matching our searches for thoroughness and consistency. The refined search strings were then applied in the selected databases to retrieve additional relevant literature.
在初始阶段，我们通过合并特定关键字（包括出版物标题、摘要和相关关键字中的术语）来完善搜索字符串。这个迭代过程使我们能够微调我们的搜索结果，并解决匹配我们的搜索的彻底性和一致性的任何潜在限制。然后将精炼的搜索字符串应用到选定的数据库中以检索其他相关文献。

In addition to our systematic search strategy, we conducted manual searches in esteemed journals and established conferences that are highly relevant to our research discipline. These journals and conferences encompass a wide range of topics, including AI, Neural Networks, Recommendation Systems, and related advancements in the relevant disciplines. By including earlier published research from these sources, we aimed to ensure comprehensive coverage of the existing literature and gather valuable insights from the forefront of the field.
除了我们的系统检索策略之外，我们还在与我们的研究学科高度相关的知名期刊和建立的会议中进行了手动检索。这些期刊和会议涵盖广泛的主题，包括人工智能、神经网络、推荐系统以及相关学科的相关进展。通过纳入这些来源的早期发表的研究，我们的目的是确保全面覆盖现有文献并从该领域的前沿收集有价值的见解。

In this study, we implemented a systematic search strategy to identify pertinent literature addressing filter bubble approaches in recommendation systems. To ensure comprehensive coverage, we established inclusion and exclusion criteria based on expressive and descriptive terms associated with recommender systems and the filter bubble phenomenon. These terms encompassed concepts such as "Recommender System," "Recommendation System," "Filter Bubble," "Echo Chamber," "Self Loop," and other closely related terms. By employing this approach, we aimed to tailor our search to our specific research objectives and review scope, while mitigating the potential impact of nomenclature discrepancies.Subsequently, we employed the Boolean OR operator to consolidate synonyms and related terms as part of our search strategy. This approach aimed to broaden the search scope and encompass various regions where the concept of the filter bubble has been investigated. By using the Boolean OR operator, we aimed to attain comprehensive results while avoiding redundancy. To further refine and narrow down the search outcomes, we then utilized the Boolean AND operator. This step allowed us to focus specifically on studies that concurrently addressed both recommender systems and the filter bubble phenomenon, ensuring the inclusion of relevant literature.
在这项研究中，我们实施了系统搜索策略来识别推荐系统中解决过滤气泡方法的相关文献。为了确保全面覆盖，我们根据与推荐系统和过滤气泡现象相关的表达性和描述性术语建立了纳入和排除标准。这些术语包含“推荐系统”、“推荐系统”、“过滤气泡”、“回声室”、“自循环”等概念以及其他密切相关的术语。通过采用这种方法，我们的目标是根据我们的具体研究目标和审查范围调整我们的搜索，同时减轻命名差异的潜在影响。随后，我们使用布尔 OR 运算符来合并同义词和相关术语，作为我们搜索策略的一部分。这种方法旨在扩大搜索范围并涵盖已研究过滤气泡概念的各个区域。通过使用布尔 OR 运算符，我们的目标是获得全面的结果，同时避免冗余。为了进一步细化和缩小搜索结果，我们使用了布尔 AND 运算符。这一步使我们能够专门关注同时解决推荐系统和过滤气泡现象的研究，确保包含相关文献。

During our systematic review, we initially collected 312 papers. To ensure the relevance of the literature, we applied basic criteria such as title, abstract, and topic alignment with our research question. We then established detailed inclusion and exclusion criteria to streamline the selection process. The inclusion criteria encompassed papers proposing solutions, addressing the existence of the filter bubble, implementing techniques, or proposing enhanced versions of recommender systems to mitigate its effects. Conversely, the exclusion criteria were applied to exclude publications that did not specifically address the filter bubble, focused on other applications or research sectors, or compared various recommendation techniques. One author took the lead in the selection strategy and conducted the initial screening, ensuring consistency with our research theme. Any disagreements regarding the suitability of specific works were resolved through discussions with other authors. After removing duplicates, we identified 185 unique articles. We then conducted a thorough assessment of the remaining articles by carefully reviewing titles, abstracts, and conclusions. Based on this assessment, we narrowed down the selection to 55 articles that exhibited relevance based on title and abstract. In the subsequent stage, we applied the specified inclusion and exclusion criteria to the remaining articles, leading to the exclusion of certain studies that did not meet our criteria.
在我们的系统审查过程中，我们最初收集了 312 篇论文。为了确保文献的相关性，我们应用了标题、摘要和主题与我们的研究问题的一致性等基本标准。然后，我们制定了详细的纳入和排除标准，以简化选择过程。纳入标准包括提出解决方案、解决过滤气泡的存在、实施技术或提出推荐系统的增强版本以减轻其影响的论文。相反，排除标准用于排除那些没有专门解决过滤气泡、专注于其他应用或研究部门或比较各种推荐技术的出版物。一位作者牵头制定了选择策略并进行了初步筛选，确保与我们的研究主题保持一致。关于特定作品的适用性的任何分歧都通过与其他作者的讨论来解决。删除重复项后，我们确定了 185 篇独特的文章。然后，我们通过仔细审查标题、摘要和结论，对其余文章进行了彻底评估。基于此评估，我们将选择范围缩小到 55 篇根据标题和摘要表现出相关性的文章。在后续阶段，我们对其余文章应用了指定的纳入和排除标准，从而排除了某些不符合我们标准的研究。

Typically, during the selection process, we applied the following exclusion criteria to refine our literature set:
通常，在选择过程中，我们应用以下排除标准来完善我们的文献集：

• •

  Duplicate records

   • 重复记录
• •

  Papers that did not comment on the existence of a filter bubble.

  
  • 没有评论过滤气泡存在的论文。
• •

  Papers related to the implementation of applications utilizing previous RSs.

  
  • 与利用以前的RS 实施应用程序相关的论文。
• •

  Papers related to research sectors other than RSs.

  
  • 与RS 以外的研究领域相关的论文。
• •

  Papers that compared various recommendation techniques.

  
  • 比较各种推荐技术的论文。
• •

  Papers written in languages other than English.

  
  • 用英语以外的语言撰写的论文。

Furthermore, the following inclusion criteria were used to select relevant literature:
此外，还使用以下纳入标准来选择相关文献：

• •

  Proposes a solution to the issue of filter bubbles in recommender systems.

  
  • 提出了推荐系统中过滤气泡问题的解决方案。
• •

  Comments on the existence of a filter bubble.

  
  • 对过滤气泡存在的评论。
• •

  Implements a technique or method to alleviate filter bubbles.

  
  • 实施减少过滤器气泡的技术或方法。
• •

  Proposes an enhanced version of recommender systems to address the problem of filter bubbles.

  
  • 提出推荐系统的增强版本，以解决过滤气泡问题。

By applying these exclusion and inclusion criteria, we ensured that the selected articles provided insights, solutions, or advancements specifically related to the filter bubble phenomenon in RSss. This process resulted in a final selection of 28 articles that met our inclusion criteria. In order to ensure a comprehensive review, we conducted a reference scan of the selected articles, which led us to identify an additional 6 relevant papers. Consequently, a total of 34 articles were included in our systematic review on the existence of the filter bubble. Figure 3 provides an overview of our research selection criteria and the distribution of publications obtained from each database.
通过应用这些排除和纳入标准，我们确保所选文章提供了与 RSss 中的过滤气泡现象特别相关的见解、解决方案或进步。这一过程最终选出了 28 篇符合我们纳入标准的文章。为了确保全面审查，我们对所选文章进行了参考扫描，从而确定了另外 6 篇相关论文。因此，我们对过滤气泡的存在进行系统评价，总共纳入了 34 篇文章。图 3 概述了我们的研究选择标准以及从每个数据库获得的出版物的分布。

In this section, we present the overall results of our study, which are based on the persuasive research, observed trends, comparative analysis, and analytical assessment conducted by all authors through a thorough debate and deliberation. Based on our findings, we have observed that research in the field of the filter bubble is growing. While the number of studies on the filter bubble is still relatively small due to its emerging nature, there has been a significant increase in research activity in recent years. As depicted in Figure 5, which illustrates the annual distribution of filter bubble studies, there were only 8 publications from 2012 to 2018, whereas in 2021 alone, there were 9 publications on the topic. Through various methodologies and datasets, the presence of a filter bubble in recommendation systems has been convincingly demonstrated. The studies have examined contextual biases using diverse datasets and platforms. Furthermore, the majority of investigations successfully illustrated the personalized effect of recommendation systems. Therefore, based on the literature we reviewed, we can confidently conclude that the filter bubble exists in recommendation systems.
在本节中，我们介绍了我们研究的总体结果，这些结果基于所有作者通过彻底的辩论和审议进行的有说服力的研究、观察到的趋势、比较分析和分析评估。根据我们的发现，我们观察到滤泡领域的研究正在不断增长。虽然由于滤泡的新兴性质，有关滤泡的研究数量仍然相对较少，但近年来研究活动显着增加。如图 5 所示，它展示了过滤气泡研究的年度分布，从 2012 年到 2018 年，只有 8 篇出版物，而仅 2021 年，就​​有 9 篇关于该主题的出版物。通过各种方法和数据集，推荐系统中过滤气泡的存在已得到令人信服的证明。这些研究使用不同的数据集和平台检查了情境偏差。此外，大多数调查成功地说明了推荐系统的个性化效果。因此，根据我们回顾的文献，我们可以自信地得出结论：推荐系统中存在过滤气泡。

The literature extensively examines various forms of bias that contribute to the problem of personalization in recommendation systems (RSs). Biases can arise at different stages, including during system design and implementation, evaluation, and user interaction. These biases can significantly impact the information gathered for system improvement and customization [90]. One prominent form of bias is algorithmic bias, which refers to biases introduced during the design and implementation of the RS. This bias can be a result of the underlying algorithms and data processing techniques used in the system. Additionally, biases can arise from the evaluation process, where researchers may unknowingly introduce their own biases into the assessment of the system’s performance. The design of the user interaction is also critical, as it can introduce additional biases in the form of presentation or exposure bias [90]. Furthermore, cognitive biases, such as confirmation bias and other behavioral biases, can influence the user’s interactions with the system and introduce biases into the data collected. These biases can affect the feedback loops used by RSs, as they are based on implicit user feedback, such as clicks or other trackable user activities. However, due to the limitations of these feedback mechanisms, the interactions are skewed towards the options presented by the system, leading to a form of bias known as presentation or exposure bias [90]. According to the research, the major causes of filter bubbles in recommendation systems can be attributed to algorithmic bias, data bias, and cognitive bias. These biases can have significant implications for the personalization and customization of RSs, and addressing them is crucial to mitigate the formation of filter bubbles.
文献广泛研究了导致推荐系统（RS）个性化问题的各种形式的偏见。偏见可能出现在不同的阶段，包括系统设计和实现、评估和用户交互期间。这些偏差会显着影响为系统改进和定制而收集的信息[90]。偏差的一种突出形式是算法偏差，它是指在 RS 的设计和实现过程中引入的偏差。这种偏差可能是系统中使用的底层算法和数据处理技术造成的。此外，评估过程中可能会产生偏差，研究人员可能会在不知不觉中将自己的偏差引入到系统性能的评估中。用户交互的设计也很重要，因为它可能会以呈现或曝光偏差的形式引入额外的偏差[90]。此外，认知偏差，例如确认偏差和其他行为偏差，可能会影响用户与系统的交互，并在收集的数据中引入偏差。这些偏差可能会影响 RS 使用的反馈循环，因为它们基于隐式用户反馈，例如点击或其他可跟踪的用户活动。然而，由于这些反馈机制的局限性，交互偏向于系统呈现的选项，导致一种称为呈现或暴露偏差的偏差[90]。研究表明，推荐系统中出现过滤气泡的主要原因可归结为算法偏差、数据偏差和认知偏差。这些偏差可能会对 RS 的个性化和定制产生重大影响，解决这些偏差对于减少过滤气泡的形成至关重要。

SSeveral research studies in the literature have proposed strategies to understand, avoid, and mitigate the harmful effects of the filter bubble phenomenon (refer to Table III and IV). This category of research explores novel ideas and diverse perspectives on how to identify and counteract the negative impact of recommendation algorithms that contribute to the formation of filter bubbles. Different approaches have been employed to determine the existence of a filter bubble, with studies utilizing benchmark datasets such as MovieLens, Twitter, or self-generated datasets. For instance, [59] conducted their research using a user interaction dataset from a WebTV platform and demonstrated that contextual bias leads to biased program recommendations, resulting in users being trapped in a filter bubble. To address this, they leveraged the Twitter social stream as an external context source, expanding the selection to include content related to social media events. They investigated the Twitter histories of key programs using two trend indicators: Trend Momentum and SigniScore. The analysis showed that Trend Momentum outperformed SigniScore, accurately predicting 96 percent of all peaks in the selected candidate program titles ahead of time. While many studies rely on datasets to support their research, some propose frameworks or models without utilizing specific datasets. For example, [82] proposed a generic framework to prevent polarization by ensuring that each user is presented with a balanced selection of content. They demonstrated how modifying a basic bandit algorithm can improve the regret bound above the state-of-the-art while satisfying the requirements for reducing polarization. These research studies offer valuable insights and methodologies for understanding and addressing the filter bubble phenomenon, providing a foundation for developing effective strategies to mitigate its negative effects in recommendation systems.
文献中的几项研究提出了理解、避免和减轻滤泡现象有害影响的策略（参见表 III 和 IV）。此类研究探讨了如何识别和抵消推荐算法导致过滤气泡形成的负面影响的新颖想法和不同观点。人们采用了不同的方法来确定过滤气泡的存在，研究利用了基准数据集，例如 MovieLens、Twitter 或自行生成的数据集。例如，[59]使用来自网络电视平台的用户交互数据集进行了研究，并证明上下文偏差会导致有偏见的节目推荐，从而导致用户陷入过滤气泡中。为了解决这个问题，他们利用 Twitter 社交流作为外部上下文源，扩大选择范围以包括与社交媒体事件相关的内容。他们使用两个趋势指标调查了关键项目的 Twitter 历史：趋势动量和 SigniScore。分析显示，Trend Momentum 的表现优于 SigniScore，提前准确预测了所选候选节目中 96% 的峰值。虽然许多研究依赖数据集来支持他们的研究，但有些研究提出的框架或模型没有利用特定的数据集。例如，[82]提出了一个通用框架，通过确保向每个用户提供平衡的内容选择来防止两极分化。他们演示了如何修改基本老虎机算法可以将后悔界限提高到最先进水平之上，同时满足减少极化的要求。 这些研究为理解和解决过滤气泡现象提供了宝贵的见解和方法，为制定有效策略以减轻其在推荐系统中的负面影响奠定了基础。

Examining users’ behavior is another important aspect of identifying the filter bubble phenomenon. For instance, [59] incorporated the Twitter social stream as an external context source to expand the selection of items to include those related to social media events. They recognized the significance of users’ behavior in determining the composition of the filter bubble. Similarly, [57] investigated the biases of four algorithms based on five metrics (relevance, variety, novelty, unexpectedness, and serendipity) across user groups categorized by eight different characteristics. To gain insight into the identified biases, they analyzed users’ behavioral patterns, such as their inclination to provide more favorable ratings. The study found that biases varied to a greater extent among user groups based on their age and curiosity levels. Despite the range of research projects conducted in this area, there is a common observation that real-time implementation of the proposed methodologies in recommendation systems has received limited attention. The practical application and integration of these research findings into real-world recommendation systems have been identified as an important area for future exploration and development.
检查用户的行为是识别过滤气泡现象的另一个重要方面。例如，[59]将 Twitter 社交流合并为外部上下文源，以扩展项目的选择以包括与社交媒体事件相关的项目。他们认识到用户行为在确定过滤气泡成分方面的重要性。同样，[57]基于五个指标（相关性、多样性、新颖性、意外性和偶然性）调查了四种算法在按八种不同特征分类的用户组中的偏差。为了深入了解已发现的偏见，他们分析了用户的行为模式，例如他们提供更有利评级的倾向。研究发现，根据年龄和好奇心水平，用户群体之间的偏见差异更大。尽管在该领域开展了一系列研究项目，但普遍认为，推荐系统中所提出的方法的实时实施受到的关注有限。这些研究成果的实际应用以及将其整合到现实世界的推荐系统中已被确定为未来探索和发展的重要领域。

Graph/network-based analysis and visualization have been employed by researchers to investigate the presence of the filter bubble. For instance, [65] developed FRediECH, a system that combines echo chamber awareness with user representations to balance the relevance, diversity, and originality of friend suggestions. FRediECH utilizes a Deep Wide architecture and a graph convolutional network to enhance the diversity of recommendations by re-ranking the results based on the network’s explicit community structure. However, this approach may have limitations as it requires defining the criteria for identifying such groups. FRediECH aims to adapt the community structure to changes in user interactions and content patterns, striking a balance between relevance and variety. In another study, [68] employed a CNN-based deep neural network technique to construct article embeddings for news articles using information such as article title, synopsis, full text, and tags from datasets. They utilized the Maximal Marginal Relevance (MMR) re-ranking technique, which compares the results of the suggested approaches with a diversified baseline. The MMR-based method evaluates multiple performance criteria, such as accuracy and variety, to re-rank items from the original recommendation list. While MMR-based methods help reduce the impact of the filter bubble, they are often criticized for being computationally expensive and sacrificing relevance for diversity, making them less feasible in real-world scenarios. Addressing these concerns, [18] proposed a novel approach called Targeted Diversification VAE-based Collaborative Filtering (TD-VAE-CF) to mitigate political polarization in media recommendations. This approach aims to strike a balance between relevance and diversity by leveraging the capabilities of Variational Autoencoders (VAE) in generating diverse and targeted recommendations.
研究人员已采用基于图形/网络的分析和可视化来研究过滤气泡的存在。例如，[65]开发了FRediECH，这是一个将回声室意识与用户表示相结合的系统，以平衡好友建议的相关性、多样性和原创性。 FRediECH 利用 Deep Wide 架构和图卷积网络，根据网络的显式社区结构对结果进行重新排序，从而增强推荐的多样性。然而，这种方法可能有局限性，因为它需要定义识别此类群体的标准。 FRediECH 旨在使社区结构适应用户交互和内容模式的变化，在相关性和多样性之间取得平衡。在另一项研究中，[68] 采用基于 CNN 的深度神经网络技术，使用数据集中的文章标题、概要、全文和标签等信息构建新闻文章的文章嵌入。他们利用最大边际相关性（MMR）重新排序技术，将建议方法的结果与多样化的基线进行比较。基于 MMR 的方法评估多个性能标准，例如准确性和多样性，以对原始推荐列表中的项目进行重新排序。虽然基于 MMR 的方法有助于减少过滤气泡的影响，但它们经常被批评为计算成本昂贵并且牺牲了多样性的相关性，使得它们在现实场景中不太可行。为了解决这些问题，[18]提出了一种称为基于 VAE 的协同过滤（TD-VAE-CF）的新颖方法，以减轻媒体推荐中的政治极化。 这种方法旨在通过利用变分自动编码器（VAE）生成多样化和有针对性的推荐的能力，在相关性和多样性之间取得平衡。

After identifying the presence of filter bubbles, many studies have proposed potential solutions. The first category of solutions focuses on bypassing or modifying algorithms. In our selected research, a significant number of solutions concentrated on enhancing content diversity. For instance, [68] and [75] presented scenarios to make session-based recommendation systems more diversity-aware by considering not only a user’s current session interactions but also diverse content from other sessions. Additionally, [82] proposed a flexible framework that allows users to have control over the source from which recommendations are selected, thereby reducing polarization in personalized systems. Furthermore, some researchers have identified strategies to enable users to explore fresh information that was previously unknown to them ([65]). To achieve content diversity, the two most commonly used approaches in recommendation systems are re-ranking and diversity modeling. Re-ranking methods, such as those proposed by [65], [68], and [73], involve post-processing techniques that reorder the ranked list provided by the baseline recommender. They assess the diversity of suggestions on the candidate list and perform a re-ranking based on this criterion. While these strategies can enhance diversity, they often require additional post-processing steps and can be computationally expensive. On the other hand, diversity modeling approaches, as suggested by [64], [70], and [77], involve modifying the core algorithm itself to make it more diversity-aware. These approaches adapt the recommendation algorithm to incorporate diversity as a key consideration (see Table V).
在确定过滤气泡的存在后，许多研究提出了潜在的解决方案。第一类解决方案侧重于绕过或修改算法。在我们选择的研究中，大量解决方案集中于增强内容多样性。例如，[68]和[75]提出了一些场景，通过不仅考虑用户当前的会话交互，还考虑其他会话的不同内容，使基于会话的推荐系统更具多样性意识。此外，[82]提出了一个灵活的框架，允许用户控制选择推荐的来源，从而减少个性化系统中的两极分化。此外，一些研究人员已经确定了使用户能够探索他们以前未知的新信息的策略（[65]）。为了实现内容多样性，推荐系统中最常用的两种方法是重新排序和多样性建模。重新排序方法，例如[65]、[68]和[73]提出的方法，涉及对基线推荐器提供的排序列表进行重新排序的后处理技术。他们评估候选列表上建议的多样性，并根据此标准进行重新排名。虽然这些策略可以增强多样性，但它们通常需要额外的后处理步骤，并且计算成本可能很高。另一方面，如[64]、[70]和[77]所建议的多样性建模方法涉及修改核心算法本身以使其更具多样性意识。这些方法调整推荐算法，将多样性作为关键考虑因素（见表 V）。

Several researchers have explored the incorporation of diversity regularization into matrix factorization (MF) models to achieve multi-objective recommendations that maximize both accuracy and variety. In their study, [78] utilize a probabilistic matrix factorization approach ([91]) to predict ratings, which has shown significant success in terms of prediction accuracy and scalability. Similarly, [81] propose two models, namely popularity-based and distance-based novelty-aware MF, which allow for a trade-off between matrix factorization performance and the requirement for novelty while only moderately sacrificing accuracy. The results of their experiments suggest that it is possible to achieve high accuracy while also introducing unique and diverse recommendations.
一些研究人员探索了将多样性正则化纳入矩阵分解（MF）模型，以实现多目标推荐，从而最大限度地提高准确性和多样性。在他们的研究中，[78]利用概率矩阵分解方法（[91]）来预测评级，该方法在预测准确性和可扩展性方面取得了显着的成功。类似地，[81]提出了两种模型，即基于流行度的模型和基于距离的新颖性感知 MF，它们允许在矩阵分解性能和新颖性要求之间进行权衡，同时仅适度牺牲准确性。他们的实验结果表明，可以实现高精度，同时引入独特且多样化的推荐。

In summary, the majority of research in this area focuses on enhancing diversity in recommendations while still maintaining a level of personalization. Additionally, there is a strong emphasis on making the recommendation process more transparent and explainable, as well as involving users in the decision-making process. Many researchers have also highlighted the importance of developing frameworks or models that are efficient and feasible for real-world scenarios. Building upon these insights, the authors of this study propose generalized methods to mitigate the filter bubble phenomenon in recommender systems, which will be discussed in the next section.
总之，该领域的大多数研究重点是增强推荐的多样性，同时仍保持一定程度的个性化。此外，还非常重视使推荐过程更加透明和可解释，以及让用户参与决策过程。许多研究人员还强调了开发针对现实场景有效且可行的框架或模型的重要性。基于这些见解，本研究的作者提出了减轻推荐系统中过滤气泡现象的通用方法，这将在下一节中讨论。

We know that the filter bubble is created due to highly personalized recommendations. A possible way to avoid this situation is to add some diversity to the recommendations through various means, including random recommendations. However, we can not completely neglect the personalized recommendations generated through previous user experiences. The solution lies in a balance between personalized and diversified recommendations. Both components are necessary but of competing nature, i.e., increasing one will decrease the other. Such conflict situations can be seen and modeled as a multi-objective optimization problem. The solution to a multi-objective optimization problem is a set of ‘non-inferior’ or ‘non-dominated’ solutions called a Pareto-optimal front.
我们知道过滤气泡是由于高度个性化的推荐而产生的。避免这种情况的一种可能方法是通过各种手段为推荐添加一些多样性，包括随机推荐。然而，我们不能完全忽视通过以往的用户体验产生的个性化推荐。解决方案在于个性化推荐和多元化推荐之间的平衡。这两个组成部分都是必要的，但具有竞争性，即增加一个组成部分会减少另一个组成部分。这种冲突情况可以被视为多目标优化问题并被建模。多目标优化问题的解决方案是一组“非劣质”或“非支配”解决方案，称为帕累托最优前沿。

Theoretically, this set contains infinitely many points for which no solution can be said better than the others. For example, a possible solution Pareto set for filter bubble could be: 100% personalization, 0% diversification, 90% personalization, 10% diversification,…, 50% personalization, 50% diversification,…, 0% personalization, 100% diversification. The first solution of the solution set 100% personalization, 0% diversification focuses only on personalized recommendations. On the other hand, the last solution 0% personalization, 100% diversification prefers diverse recommendations only. However, there are many intermediate solutions that try to make a balance between both. An important point to note here is that one solution is not better than any other solution because each has a better value for exactly one objective. The concept of the Pareto optimal set is described in Figure 7.
理论上，这个集合包含无限多个点，对于这些点，没有任何解决方案比其他解决方案更好。例如，过滤气泡的帕累托集可能的解决方案可以是：100%个性化，0%多样化，90%个性化，10%多样化，……，50%个性化，50%多样化，……，0%个性化，100%多样化。该解决方案的第一个方案集100%个性化、0%多样化，只注重个性化推荐。另一方面，最后一个解决方案 0% 个性化，100% 多样化更喜欢仅多样化的推荐。然而，有许多中间解决方案试图在两者之间取得平衡。这里需要注意的重要一点是，一种解决方案并不比任何其他解决方案更好，因为每种解决方案对于一个目标都有更好的价值。 Pareto 最优集的概念如图 7 所示。

The filter bubble is a result of highly personalized recommendations. To avoid this situation, it is necessary to introduce diversity into the recommendations, which can be achieved through various means, including random recommendations. However, personalized recommendations based on previous user experiences cannot be completely disregarded. The solution lies in finding a balance between personalized and diversified recommendations, recognizing that both components are necessary but inherently compete with each other. This conflict can be formulated and modeled as a multi-objective optimization problem.
过滤气泡是高度个性化推荐的结果。为了避免这种情况，需要在推荐中引入多样性，这可以通过多种手段来实现，包括随机推荐。然而，基于以往用户体验的个性化推荐也不能完全忽视。解决方案在于在个性化和多样化推荐之间找到平衡，认识到这两个组成部分都是必要的，但本质上是相互竞争的。这种冲突可以表述为多目标优化问题并建模。

In a multi-objective optimization problem, the solution space consists of a set of ’non-inferior’ or ’non-dominated’ solutions known as the Pareto-optimal front. Theoretically, this set comprises infinitely many points, with no solution being considered better than others. For instance, in the context of addressing the filter bubble, the Pareto set may include solutions such as 100% personalization, 0% diversification, 90% personalization, 10% diversification, …, 50% personalization, 50% diversification, …, 0% personalization, 100% diversification. The first solution in the set, 100% personalization, 0% diversification, focuses solely on personalized recommendations, while the last solution, 0% personalization, 100% diversification, prioritizes diverse recommendations. However, there exist many intermediate solutions that aim to strike a balance between both objectives. It is important to note that no single solution is superior to others since each solution offers better values for a specific objective. The concept of the Pareto-optimal set is illustrated in Figure 7.
在多目标优化问题中，解空间由一组“非劣质”或“非支配”解组成，称为帕累托最优前沿。理论上，该集合包含无限多个点，没有一个解决方案被认为比其他解决方案更好。例如，在解决过滤泡沫的背景下，帕累托集可能包括诸如 100% 个性化、0% 多元化、90% 个性化、10% 多元化、…、50% 个性化、50% 多元化、…、0% 等解决方案个性化，100%多样化。该组中的第一个解决方案，100% 个性化，0% 多样化，仅关注个性化推荐，而最后一个解决方案，0% 个性化，100% 多样化，优先考虑多样化推荐。然而，存在许多旨在在两个目标之间取得平衡的中间解决方案。值得注意的是，没有任何一种解决方案优于其他解决方案，因为每种解决方案都为特定目标提供了更好的价值。帕累托最优集的概念如图 7 所示。

Here, the solutions A, B, and C are incomparable but all of them are better than solution D and E. If the filter bubble problem is posed as a bi-objective optimization problem, it may be represented as Eq. 1:
这里，解A、B和C没有可比性，但都优于解D和E。如果将滤泡问题提出为双目标优化问题，则可以将其表示为式(1)。 1：

The Diversity Score measures the degree of diversified recommendations, while the Personalization Score represents the degree of personalized recommendations in the final outcome, both normalized to the range [0,1]. The Pareto set of Eq. 1 is depicted in Figure 8. In this figure, point P (0,1) represents a solution that emphasizes full personalization, while point D (1,0) represents a completely random recommendation. Recommendations A, B, and C fall within the Desirable Area of the Pareto-optimal front, exhibiting non-zero values for both scores, but with varying degrees. Recommendation A contains more personalized information than B and C, while C has a higher level of diversity. Once we have developed such a theoretical model, the next step is to define the mathematical formulation of Eq. 1, which involves determining the formulas for calculating the Diversity Score and the Personalization Score. By solving Eq. 1, we can obtain a set of recommendations that have incomparable values of personalization and diversity scores. Recommendations falling within the Desirable Area are expected to generate bubble-free results.
多样性得分衡量推荐的多样化程度，而个性化得分则代表最终结果中个性化推荐的程度，两者均标准化为范围 [0,1]。帕累托方程组。图8描绘了图1。在该图中，点P(0,1)代表强调完全个性化的解决方案，而点D(1,0)代表完全随机推荐。建议 A、B 和 C 属于帕累托最优前沿的理想区域，两个分数都呈现非零值，但程度不同。推荐A比B和C包含更多的个性化信息，而C则具有更高水平的多样性。一旦我们建立了这样的理论模型，下一步就是定义方程的数学公式。 1、确定多样性得分和个性化得分的计算公式。通过求解方程。 1，我们可以获得一组具有无与伦比的个性化和多样性分数价值的推荐。落在理想区域内的建议预计会产生无泡沫的结果。

Based on the insights gained from our research, we propose an architecture for integrated tools that can be employed in recommendation systems to mitigate the formation of filter bubbles. Drawing upon the findings of our literature analysis, we suggest that this integrated tool should serve two primary functions: (1) alerting users to the potential presence of a filter bubble, and (2) allowing users to customize the extent of personalization.
根据我们研究中获得的见解，我们提出了一种集成工具架构，可用于推荐系统，以减少过滤气泡的形成。根据我们的文献分析结果，我们建议这个集成工具应具有两个主要功能：（1）提醒用户潜在存在过滤气泡，（2）允许用户自定义个性化的程度。

In recent times, there has been a growing interest in explainable artificial intelligence (XAI) across various research domains, aiming to address the challenges posed by increasing complexity, scalability, and automation [4, 92]. Consequently, the development of explainable recommendation systems (XRSs) has gained momentum. Notably, researchers such as Peake et al. [93] have proposed a novel approach for extracting explanations from latent factor recommendation systems by employing training association rules on the outcomes of a matrix factorization black-box model. Their method effectively balances interpretability and accuracy without compromising flexibility or relying on external data sources. Explanations play a crucial role in ensuring that users comprehend and trust recommendation systems that prioritize explainability. Without accompanying explanations, there is a risk that the recommendations generated by a system may be perceived as untrustworthy or lacking authenticity [94]. By understanding the rationale behind a recommendation, users can identify potential filter bubbles and take steps to burst them. For instance, if an item is accompanied by a rating indicating the level of personalization in the suggestion, whether it is based on previous searches or purely random [16], users can gain insights into why the recommendation is being made.
近年来，人们对各个研究领域的可解释人工智能 (XAI) 越来越感兴趣，旨在解决日益复杂性、可扩展性和自动化带来的挑战 [4, 92]。因此，可解释推荐系统（XRS）的发展势头强劲。值得注意的是，皮克等人的研究人员。 [93]提出了一种新方法，通过对矩阵分解黑盒模型的结果采用训练关联规则，从潜在因素推荐系统中提取解释。他们的方法有效地平衡了可解释性和准确性，而不影响灵活性或依赖外部数据源。解释在确保用户理解和信任优先考虑可解释性的推荐系统方面发挥着至关重要的作用。如果没有附带解释，系统生成的建议可能会被认为不可信或缺乏真实性[94]。通过了解推荐背后的基本原理，用户可以识别潜在的过滤气泡并采取措施打破它们。例如，如果一个项目附有一个评级，表明建议中的个性化程度，无论是基于之前的搜索还是纯粹随机的[16]，用户都可以深入了解为什么提出建议。

In line with designing a fair and explainable system, an XRS focused on food recipe recommendations has been proposed [95]. The notable contribution of this recommendation approach is its comprehensive inclusion of explainability features, which not only provide explanations for recommendations but also raise nutrition awareness. By incorporating additional aspects into the explanation process, this approach aims to enhance user satisfaction and understanding, making it a valuable component of an XRS.
为了设计一个公平且可解释的系统，人们提出了一种专注于食物食谱推荐的 XRS [95]。这种推荐方法的显着贡献是它全面包含了可解释性特征，不仅为推荐提供了解释，而且提高了营养意识。通过将其他方面纳入解释过程，该方法旨在提高用户满意度和理解力，使其成为 XRS 的重要组成部分。

Balancing the trade-off between personalization and diversification is crucial when recommending items in order to address the filter bubble phenomenon. Customized recommendations are important as they facilitate the user’s search for relevant items. However, it is equally important to provide diverse results to break the bubble effect. Therefore, we aim to incorporate this trade-off into our tools and give users the ability to choose the type of recommendations they desire. By providing users with control over this trade-off, the recommendation system can achieve its goal while also preventing users from being trapped in a filter bubble. For instance, if a user prefers items that are similar to their previous searches, the degree of personalization can be adjusted to provide more tailored recommendations. On the other hand, if a user wants to explore a wider range of items without being influenced by their past data, they can modify the degree of personalization to receive more diverse recommendations.
在推荐商品以解决过滤泡沫现象时，平衡个性化和多样化之间的权衡至关重要。定制推荐很重要，因为它们有助于用户搜索相关项目。然而，提供多样化的结果来打破泡沫效应也同样重要。因此，我们的目标是将这种权衡纳入我们的工具中，并使用户能够选择他们想要的推荐类型。通过为用户提供对这种权衡的控制，推荐系统可以实现其目标，同时还可以防止用户陷入过滤气泡中。例如，如果用户更喜欢与其之前搜索相似的项目，则可以调整个性化程度以提供更量身定制的推荐。另一方面，如果用户想要探索更广泛的项目而不受到过去数据的影响，他们可以修改个性化程度以接收更多样化的推荐。

The tool proposed in Figure 9 aims to provide users with a better understanding of the recommendations they receive and empower them to customize their future searches to break free from the filter bubble. By offering transparency and explanation, users can gain insights into why a specific recommendation was made, allowing them to make informed decisions and challenge the bubble effect. Figure 9 also illustrates a comparison between the proposed explainable recommendation system and a standard recommendation system. The added layer of explainability in the proposed system enhances user understanding and trust, promoting a more satisfying user experience. Figure 10 focuses on the tool’s interface, using a movie suggestion example. In this scenario, the user’s preferences primarily revolve around action, thriller, and drama movies, as depicted in the figure. When the system is personalized, the user is presented with recommendations that align with their preferred genres. On the other hand, when the degree of personalization is adjusted towards diversity, the system recommends a broader range of content, allowing the user to explore movies beyond their usual preferences.
图 9 中提出的工具旨在让用户更好地理解他们收到的推荐，并使他们能够定制未来的搜索以摆脱过滤泡沫。通过提供透明度和解释，用户可以深入了解为什么提出特定建议，从而使他们能够做出明智的决策并挑战泡沫效应。图 9 还说明了所提出的可解释推荐系统和标准推荐系统之间的比较。所提出的系统中增加的可解释性层增强了用户的理解和信任，促进更令人满意的用户体验。图 10 使用电影建议示例重点介绍了该工具的界面。在这种场景下，用户的偏好主要围绕动作片、惊悚片和剧情片，如图所示。当系统个性化时，会向用户呈现与其偏好类型相符的推荐。另一方面，当个性化程度向多样性调整时，系统会推荐更广泛的内容，让用户探索超出其平时偏好的电影。

As discussed in previous sections, the primary solution to combat the filter bubble problem is to incorporate diverse content in recommendations. However, it is crucial to define diversity itself as it encompasses various types, each with its specific definition and implications. It is worth noting that current recommendation systems intentionally introduce some level of variety to ensure that the recommended items are not excessively similar [96]. Additionally, other types of diversity, such as personalized and temporal diversity, are also being utilized in recommendation systems [97]. While measures of diversity are already employed in recommendation systems, their objective has not always been to address the filter bubble issue but rather to provide users with a range of somewhat dissimilar options to choose from. Consequently, it becomes crucial to define diversity in the context of the filter bubble phenomenon. In selecting an appropriate diversity measure, several key considerations should be taken into account.
正如前面几节所讨论的，解决过滤气泡问题的主要解决方案是将不同的内容纳入推荐中。然而，定义多样性本身至关重要，因为它包含各种类型，每种类型都有其特定的定义和含义。值得注意的是，当前的推荐系统有意引入一定程度的多样性，以确保推荐的项目不会过度相似[96]。此外，其他类型的多样性，例如个性化和时间多样性，也被用于推荐系统[97]。虽然推荐系统中已经采用了多样性衡量标准，但其目标并不总是解决过滤气泡问题，而是为用户提供一系列有些不同的选项供选择。因此，在过滤气泡现象的背景下定义多样性变得至关重要。在选择适当的多样性措施时，应考虑几个关键因素。

• •

  Opposite of similarity: In early recommendation systems, diversity was viewed as the opposite of similarity and defined as (1 - similarity), where similarity is a measure of the proximity between user interests and recommended items [96]. In a list of items, diversity is calculated as the average dissimilarity between all pairs of items.

  
  • 相似性的反面：在早期的推荐系统中，多样性被视为相似性的反面，并定义为（1 - 相似性），其中相似性是用户兴趣和推荐项目之间的接近程度的度量[96]。在项目列表中，多样性计算为所有项目对之间的平均差异。
• •

  Diversity through Rearrangement/Re-ranking: This approach involves rearranging the list of recommended items generated by the algorithm to improve the diversity metric [98]. It has been observed that this simple approach works well in certain scenarios. It can be seen as an optimization problem that aims to maximize the diversity metric.

  
  • 通过重新排列/重新排名来实现多样性：这种方法涉及重新排列算法生成的推荐项目列表，以改进多样性度量[98]。据观察，这种简单的方法在某些情况下效果很好。它可以被视为一个旨在最大化多样性度量的优化问题。
• •

  Diversity in items and/or source: It is important to decide whether diversity should be introduced only in the recommended content or in the content provider as well [99]. For example, in online shopping, diversified items may include different garments, while diversified sources may involve different brands.

  
  • 项目和/或来源的多样性：重要的是决定是否应仅在推荐内容中引入多样性，还是在内容提供商中也引入多样性[99]。例如，在网上购物中，多元化的商品可能包括不同的服装，而多元化的来源可能涉及不同的品牌。
• •

  Personalized/User-specific Diversity: Diversity can be introduced irrespective of user profiles, which is referred to as non-personalized diversity. However, it is considered better to capture the diversity needs of individuals by modeling their characteristics and incorporating them into the diversity metric [100]. Such diversity measures are known as personalized matrices.

  
  • 个性化/特定于用户的多样性：可以不考虑用户简档而引入多样性，这被称为非个性化多样性。然而，通过对个体的特征进行建模并将其纳入多样性度量中来更好地捕获个体的多样性需求[100]。这种多样性度量被称为个性化矩阵。
• •

  Temporal Diversity: In certain domains, recommendations need to consider the dimension of time, giving rise to the concept of temporal diversity [101]. News recommendation systems, for instance, must account for rapidly changing news topics, as well as the evolving preferences of users over different time periods (weekly, monthly, yearly, or seasonally). Thus, temporal diversity should be designed to address users’ short- and long-term preferences.

  
  • 时间多样性：在某些领域，建议需要考虑时间维度，从而产生了时间多样性的概念[101]。例如，新闻推荐系统必须考虑快速变化的新闻主题，以及用户在不同时间段（每周、每月、每年或季节性）不断变化的偏好。因此，时间多样性的设计应满足用户的短期和长期偏好。
• •

  Hybrid Diversity: A diversity metric may incorporate multiple aspects discussed above, resulting in a hybrid diversity measure [102]. A simple implementation could involve calculating a weighted sum of various diversity measures to capture different dimensions of diversity.

  
  • 混合分集：分集度量可以包含上面讨论的多个方面，从而产生混合分集度量[102]。一个简单的实现可能涉及计算各种多样性度量的加权和以捕获不同维度的多样性。

Overall, the process of selecting the right diversity metric is a meticulous task that involves careful consideration of various factors. To ensure an effective selection, the following steps need to be followed:
总的来说，选择正确的多样性指标的过程是一项细致的任务，需要仔细考虑各种因素。为了确保有效的选择，需要遵循以下步骤：

1. 1.

   Study the specific domain of the recommendation system under consideration. This involves understanding the characteristics of the items, the preferences of the users, and any temporal or contextual factors that may influence recommendations.

   
   1. 研究所考虑的推荐系统的特定领域。这涉及了解项目的特征、用户的偏好以及可能影响推荐的任何时间或上下文因素。
2. 2.

   Define diversity in the context of the predetermined domain. This entails identifying the specific dimensions or aspects of diversity that are relevant and meaningful for the given domain.

   
   2. 在预定领域的背景下定义多样性。这需要确定与给定领域相关且有意义的多样性的具体维度或方面。
3. 3.

   Select appropriate diversity measure(s) that align with the defined notion of diversity. This may involve choosing from existing diversity metrics or developing new ones tailored to the specific requirements of the domain.

   
   3. 选择符合所定义的多样性概念的适当多样性措施。这可能涉及从现有的多样性指标中进行选择或开发适合该领域特定要求的新指标。
4. 4.

   Combine the selected diversity measure(s) with an appropriate prevention approach to effectively address the filter bubble problem. This could involve incorporating diversity constraints into recommendation algorithms or utilizing post-processing techniques for re-ranking recommendations.

   
   4. 将选定的多样性措施与适当的预防方法相结合，以有效解决过滤气泡问题。这可能涉及将多样性约束纳入推荐算法或利用后处理技术对推荐进行重新排名。
5. 5.

   Gather feedback from users, either implicitly through user interactions or explicitly through surveys or interviews, to evaluate the effectiveness of the diversity measures and their impact on user satisfaction.

   
   5. 通过用户互动隐式或通过调查或访谈明确地收集用户反馈，以评估多样性措施的有效性及其对用户满意度的影响。
6. 6.

   Adapt and modify the diversity measure(s) based on the received feedback. This iterative process ensures that the diversity metric continues to capture the evolving needs and preferences of the users.

   
   6. 根据收到的反馈调整和修改多样性措施。这个迭代过程确保多样性指标继续捕捉用户不断变化的需求和偏好。

• •

  Defining diversity in a domain-specific context: Diversity plays a critical role in addressing the filter bubble problem, but its definition may vary depending on the recommendation domain [103]. For instance, diversity in a movie recommendation system may differ from diversity in an online clothing portal. It is important to establish domain-specific definitions of diversity and develop mathematical frameworks accordingly. It is worth noting that similar concepts to diversity, such as novelty and serendipity, have been discussed in the literature [104, 105]. While diversity refers to the presence of variety in a recommended item list, novelty captures the difference between past and present recommendations, and serendipity occurs when new and relevant but previously unknown items are included in the recommendations. The choice of which concept or combination to utilize should be based on the specific requirements of the application.

  
  • 在特定领域的背景下定义多样性：多样性在解决过滤气泡问题中发挥着关键作用，但其定义可能会根据推荐领域的不同而有所不同[103]。例如，电影推荐系统中的多样性可能与在线服装门户中的多样性不同。建立特定领域的多样性定义并相应地开发数学框架非常重要。值得注意的是，文献[104, 105]中已经讨论了与多样性类似的概念，例如新颖性和偶然性。多样性是指推荐项目列表中存在多样性，而新颖性则体现了过去和当前推荐之间的差异，而当推荐中包含新的相关但以前未知的项目时，就会出现偶然性。选择使用哪种概念或组合应基于应用程序的具体要求。
• •

  Exploring contrasting recommendations or opinions: When addressing the filter bubble issue, incorporating contrasting recommendations or opinions can promote a more balanced understanding, particularly in news recommendation systems. However, it is necessary to define the concept of "Opposite Recommendations" and establish domain-specific definitions to effectively incorporate this approach. It should be noted that defining "Opposite" is relatively straightforward in domains like news recommendation but may pose challenges in other domains, such as book recommendations [106].

  
  • 探索对比的建议或意见：在解决过滤气泡问题时，纳入对比的建议或意见可以促进更平衡的理解，特别是在新闻推荐系统中。然而，有必要定义“相反建议”的概念并建立特定领域的定义以有效地纳入这种方法。应该注意的是，定义“相反”在新闻推荐等领域相对简单，但在其他领域可能会带来挑战，例如书籍推荐[106]。
• •

  Identifying sources responsible for spreading fake news: Identifying sources responsible for spreading fake news is crucial in addressing the filter bubble problem. Fake news or misinformation greatly contributes to the issue. However, developing advanced natural language processing (NLP) techniques that can effectively detect fake news poses a challenge, especially when dealing with aspects such as sarcasm and deceptive language usage. Deep learning-based NLP models like Deep Bidirectional Transformers, along with techniques like transfer learning and fine-tuning, can be explored to enhance language understanding and mitigate the negative impact of the filter bubble [107].

  
  • 识别传播假新闻的来源：识别传播假新闻的来源对于解决过滤泡沫问题至关重要。假新闻或错误信息在很大程度上加剧了这一问题。然而，开发能够有效检测假新闻的先进自然语言处理（NLP）技术是一项挑战，特别是在处理讽刺和欺骗性语言使用等方面时。可以探索基于深度学习的 NLP 模型（例如深度双向变压器）以及迁移学习和微调等技术，以增强语言理解并减轻过滤气泡的负面影响[107]。
• •

  Establishing the relationship between domain-specific external factors and the filter bubble: Establishing the relationship between domain-specific external factors and the filter bubble is crucial in understanding and addressing this phenomenon. Various external factors, such as the presence of fake news in news recommendation systems, contribute to the filter bubble problem. It is important to investigate and comprehend the connection between these factors and the filter bubble. Tracing the origins of misinformation is a vital step in addressing the filter bubble, and advanced natural language processing (NLP) techniques can greatly assist in this process. Furthermore, the impact of the filter bubble can vary significantly across different applications. For example, in a food recommendation system, a filter bubble can have detrimental effects on users’ well-being by excluding nutritious diets and promoting a particular genre of food. The findings of [95] also highlight how the filter bubble effect can introduce intentional biases when providing choices for restaurants and related domains. On the other hand, the influence of the filter bubble may be more pronounced in a video streaming platform like YouTube, while having only a marginal effect on users in a dress/outfit recommendation system for an online clothing portal. It is essential to recognize that generalized solutions may not be effective for every application, emphasizing the need for domain-specific analysis of the filter bubble. Each application requires a tailored approach and a deeper understanding of its specific dynamics to effectively mitigate the filter bubble’s impact.

  
  • 建立特定领域的外部因素和过滤气泡之间的关系：建立特定领域的外部因素和过滤气泡之间的关系对于理解和解决这一现象至关重要。各种外部因素，例如新闻推荐系统中存在假新闻，都会导致过滤泡沫问题。研究和理解这些因素与滤泡之间的联系非常重要。追踪错误信息的根源是解决过滤气泡问题的重要一步，先进的自然语言处理 (NLP) 技术可以极大地帮助这一过程。此外，滤泡的影响在不同的应用中可能会有很大差异。例如，在食物推荐系统中，过滤气泡可能会排除营养饮食并推销特定类型的食物，从而对用户的健康产生不利影响。 [95]的研究结果还强调了过滤气泡效应在为餐馆和相关领域提供选择时如何引入故意偏差。另一方面，过滤气泡的影响在 YouTube 等视频流平台中可能更为明显，而在在线服装门户的着装推荐系统中对用户的影响却很小。必须认识到，通用解决方案可能并不适用于所有应用，这强调了对滤泡进行特定领域分析的必要性。每个应用都需要量身定制的方法并对其特定动态有更深入的了解，以有效减轻滤泡的影响。
• •

  Enhancing data quality for visualization and integration: Enhancing the quality of data is of utmost importance for effective visualization and integration in the context of the filter bubble. As emphasized by [108] C. Sardianos, I. Varlamis, C. Chronis, G. Dimitrakopoulos, A. Alsalemi, Y. Himeur, F. Bensaali, and A. Amira, “The emergence of explainability of intelligent systems: Delivering explainable and personalized recommendations for energy efficiency,” International Journal of Intelligent Systems, vol. 36, no. 2, pp. 656–680, 2021. ×, researchers should dedicate efforts to explore methods that can enhance the quality of data used in this context. By improving data quality, we can ensure more reliable and accurate results in visualization and integration processes. Furthermore, it is crucial to address the issue of information cocooning that is prevalent in news recommender systems. These systems often filter out content that users may find uninteresting, resulting in a narrowing of their information exposure over a period of approximately seven days. This can have significant implications, particularly for individuals who are heavily reliant on social media platforms. It is imperative for the research community to tackle the challenge of designing evaluation mechanisms that incorporate social filtering. By doing so, we can mitigate the potential negative consequences of information cocooning and promote a more diverse and balanced information environment for users [1].

  
  • 提高可视化和集成的数据质量：提高数据质量对于过滤气泡环境中的有效可视化和集成至关重要。正如[108]所强调的，研究人员应该致力于探索可以提高在此背景下使用的数据质量的方法。通过提高数据质量，我们可以确保可视化和集成过程中的结果更加可靠和准确。此外，解决新闻推荐系统中普遍存在的信息茧化问题也至关重要。这些系统通常会过滤掉用户可能觉得不感兴趣的内容，从而在大约 7 天的时间内缩小他们的信息暴露范围。这可能会产生重大影响，特别是对于严重依赖社交媒体平台的个人而言。研究界必须应对设计纳入社会过滤的评估机制的挑战。通过这样做，我们可以减轻信息茧化的潜在负面后果，并为用户促进更加多样化和平衡的信息环境[1]。

Several promising research directions could be pursued to mitigate the filter bubble problem:
可以采取几个有前途的研究方向来减轻过滤器气泡问题：

• •

  Diversity-aware recommendations: Designing algorithms that aim to increase the diversity of recommendations can help in mitigating the filter bubble. These algorithms need to balance the trade-off between relevance and diversity [109, 110].

  
  • 多样性感知推荐：设计旨在增加推荐多样性的算法有助于减少过滤气泡。这些算法需要平衡相关性和多样性之间的权衡[109, 110]。
• •

  Serendipity in recommendations: Developing recommendation techniques that emphasize serendipity (unexpected but useful recommendations) could help users discover new, out-of-bubble content. These methods would encourage exposure to diverse and novel items that the user might not have found otherwise [111].

  
  • 推荐中的偶然性：开发强调偶然性（意外但有用的推荐）的推荐技术可以帮助用户发现新的、泡沫之外的内容。这些方法将鼓励用户接触到用户可能不会发现的多样化和新颖的项目[111]。
• •

  Explainability and transparency: Explainable AI can help users understand why a particular recommendation was made. Seeing the rationale behind the recommendations might make users more receptive to different content, reducing the filter bubble effect [112].

  
  • 可解释性和透明度：可解释的人工智能可以帮助用户理解为什么提出特定的建议。了解推荐背后的基本原理可能会让用户更容易接受不同的内容，从而减少过滤气泡效应[112]。
• •

  User-controlled recommendations: Allowing users to have more control over their recommendations, such as adjusting the degree of novelty or diversity, could also help alleviate the filter bubble problem.

  
  • 用户控制的推荐：允许用户更好地控制他们的推荐，例如调整新颖性或多样性的程度，也有助于缓解过滤气泡问题。
• •

  Cross-domain recommendations: Leveraging data from different domains can help in providing a broader range of recommendations. For example, if a user interacts with various content types (books, movies, music), these can be used to cross-pollinate recommendations across these domains [111].

  
  • 跨领域建议：利用不同领域的数据有助于提供更广泛的建议。例如，如果用户与各种内容类型（书籍、电影、音乐）进行交互，这些可以用于跨这些领域交叉传播推荐[111]。
• •

  Fairness and bias mitigation: Actively researching and implementing algorithms that take into account and mitigate biases in recommender systems can help to ensure that the system does not favour certain types of content, hence reducing the risk of a filter bubble [113, 114].

  
  • 公平性和偏见缓解：积极研究和实施考虑并缓解推荐系统中偏见的算法，有助于确保系统不偏向某些类型的内容，从而降低过滤泡沫的风险[113, 114]。
• •

  Long-term user modeling: Traditionally, recommender systems have focused on immediate rewards (clicks, purchases, etc.), leading to a filter bubble. Research into long-term user modeling can help understand the evolving needs and tastes of users, potentially aiding in delivering a more diverse set of recommendations [115].

  
  • 长期用户建模：传统上，推荐系统关注即时奖励（点击、购买等），从而导致过滤泡沫。对长期用户建模的研究可以帮助了解用户不断变化的需求和品味，可能有助于提供更多样化的建议[115]。