System Introduction Credible Recommendation System Construction and Technology
Serving at the forefront of human-centered artificial intelligence, recommender systems (RS) are widely deployed in almost every corner of the web and facilitate human decision-making processes. However, despite the enormous capabilities and potential of RS, it may also have adverse effects on users, items, producers, platforms, and even society as a whole, such as compromising user trust due to opacity, opacity for different consumers or producers Fair treatment, privacy concerns due to extensive use of user private data for personalization, etc. All of these lead to an urgent need for trusted recommender systems (TRS) to mitigate or avoid these adverse effects and risks. In this review, we will introduce related technologies for trustworthy and responsible recommendation, including but not limited to explainable recommendation, recommendation fairness, privacy-aware recommendation, recommendation robustness, user-controllable recommendation, and how these different perspectives are worthwhile Relationships in trust and responsible referrals . We hope that through this review, readers will have a comprehensive understanding of the research field, and arouse the society's attention to the importance of credible recommendation, existing research results and future research directions.
https://www.zhuanzhi.ai/paper/cb89489a00bfefdf45e3670b928beb99
Recommender systems (RS) are widely used in various systems such as e-commerce, social networks, search engines, news portals, recruitment platforms, smart assistants, smart home and smart city services, as well as healthcare and financial applications, which provide high-quality services has been recognized for its ability to bridge the gap between users and projects by providing tailored content for everyone. Recommender systems can not only help users find relevant information more efficiently, but can also directly influence human decision-making processes by providing relevant recommendations, and even shape users’ worldviews by exposing the selected content to users. In general, recommender systems are at the forefront of human-centered AI research, a bridge between humans and AI.
However, RS has its pros and cons, and it can offer both hope and risk. There is a growing concern that irresponsible use of recommendation technology may lead to counterproductive and untrustworthy issues, such as loss of user trust due to opacity, unfair treatment of different users, producers or platforms, and heavy use of user privacy Privacy concerns with data personalization, echo chambers created by the repeated reinforcement of users’ pre-existing interests due to lack of user variability—these problems continue to expand. These vulnerabilities greatly limit the development and deployment of recommendation algorithms, and may even lead to serious economic and social problems. Therefore, it is not enough to only consider recommendation accuracy when developing modern recommender systems. We also need to ensure that the model is fair, not tampered with, does not crash under different conditions, and can be understood by humans . In addition, the design and development process of RS also needs to be transparent and inclusive. Besides accuracy, all of these considerations that make recommender systems safe, accountable, and worthy of our trust are relevant to trustworthy recommender system research. Since recommender systems are an important direction of human-centered artificial intelligence research, directly involving human loops, Trustworthy Recommender Systems (TRS) have been leading the research in Trustworthy Artificial Intelligence (TAI) over the past few years , including definitions, methods, and evaluations of trustworthiness, fairness, robustness, privacy, and how humans interact with trustworthy AI systems.
Therefore, as an important example of trustworthy AI in the context of recommender system research, in this review, we introduce trustworthy recommender systems as interpretable, fair, private, robust, and controllable Trust a competent RS at the core . We believe that combining these aspects when designing recommender systems will increase the accountability of recommender systems, gain the trust of human users, and significantly contribute to the social benefits of recommender systems. Differences from existing reviews. There have been recent reviews on specific ethical issues in recommendation scenarios, such as interpretability [61, 340], bias and fairness [54, 77, 178, 225, 288, 328], privacy protection [140, 308], user controllability [142, 143], etc. These investigations successfully highlighted the importance of social responsibility in recommender systems, leading to further development of this important research direction. However, these issues are only presented in their own independent ways, and a systematic understanding of trustworthiness in recommendations and the interrelationships between various trustworthiness perspectives is necessary. The closest ones to our study are Dong et al. [82] and Mobasher et al. [207]. However, [82] only deals with user social relations, robustness, and interpretability, and [207] only discusses attack models and algorithm robustness in recommendation, and does not examine the intrinsic relationship between these concepts. In contrast, our work introduces credibility on a more comprehensive perspective, highlights the relationships between perspectives, and sheds light on open questions and future directions for exploring the intersection of perspectives.
Relationship to other trusted AI research. Due to its importance and necessity, there is a lot of discussion and debate about what Trustworthy Artificial Intelligence (TAI) means. In particular, Toreini et al. [273] studied trust in AI and summarized the properties of AI as competence, benevolence, integrity, and predictability; Varshney [280] argued that a trustworthy machine learning system should have sufficient Basic performance, reliability, human interaction, and selflessness; Liu et al. [186] consider TAI as a non-threatening or risk-free project and focus on six dimensions to achieve credibility: security and robustness, non-discrimination and fairness, explainability, privacy, accountability and auditability, and environmental well-being. In addition, in 2019, the European Union (EU) proposed the "Trustworthy AI1 Ethics Guidelines", which requires that AI systems should meet four ethical principles: respect for human autonomy, prevention of harm, explainability and fairness [6]. While the existing literature explores the space of integrity from different perspectives, several key aspects that have received the most recognition and consensus are interpretability, fairness, privacy, robustness, and controllability, which we believe are also key to TRS component.
The primary readers of this review are RS researchers, technologists, and professionals whose goal is to make recommender systems more trustworthy and accountable . On the other hand, since the recommender system is a very representative and universal human-centered AI system, the target readers of the review also include general AI researchers, practitioners, theorists, as well as the credibility, ethics and trustworthiness of AI. and public policy makers interested in regulations. The rest of the review is organized as follows: Section 2 introduces the preliminary knowledge of recommender systems and some representative recommendation algorithms. Sections 3, 4, 5, 6, and 7 focus on explainability, fairness, privacy, robustness, and controllability, respectively. The last part is a summary of the full text.
interpretability
Explainable recommendation has been an important area in industry and academia, aiming to improve the transparency, user satisfaction and trustworthiness of recommender systems [340, 341]. Specifically, the goal is to provide understandable reasons along with recommended items to help stakeholders make better and reliable decisions, while increasing the transparency and credibility of recommender systems. Explanations in recommender systems can help model developers to understand and debug how the decision-making process works, and can also facilitate better engagement and trust by end users using the results produced by the system. In addition to web portal-based recommender systems, explanations are also integrated into conversational recommender interfaces, such as Apple Siri, Microsoft Cortana, and Amazon Alexa, as direct interaction portals for end users [62]. They are able to provide clear user preferences, intelligently interact with users, and provide explanations in conjunction with certain suggestions.
fairness
Recommender systems have long been considered “goodwill” systems that help users (e.g., by helping them find relevant items) and create value for businesses (e.g., increase sales or improve customer retention) [141] . In recent years, however, considerable attention has been paid to the issue of fairness in proposals from both academia and industry [178]. Several studies suggest that RS may be vulnerable to inequity in several ways, which may have adverse consequences for underrepresented or disadvantaged groups [109,176,180,254]. For example, in an e-commerce system, RS may primarily promote profit maximization for certain producers [103], or in an online job market, RS may lead to racial or gender discrimination, disproportionately recommending low rates to certain groups of users Paid work [109]. Therefore, in order to improve the satisfaction of different stakeholders in RS[2], it is important to establish a credible and responsible system for fairness in research recommendation.
privacy
With the growing focus on machine learning methods for collecting and analyzing personal data, the ethical need for data privacy has been formally recognized in mandatory regulations and laws [22, 281]. Therefore, research on privacy-preserving machine learning has grown significantly in recent years [185]. It is believed that a more trustworthy web service should provide privacy-preserving solutions to avoid unwanted exposure of information by any participant of the system. In both recommender systems and machine learning in general, multiple definitions of privacy exist [5, 148, 255, 257], and in most cases they have the same components:
-
Private Information: Critical or valuable information that requires restricted access. For example, user identity and user-sensitive attributes (such as gender, age, address, etc.).
-
Ownership: Only authorized entities can access and control the corresponding private information, which may refer to users or even the platform itself.
-
Threat: Malicious entities (internal or external to the system) designed to obtain or manipulate private information. Note that these entities may exploit secondary public information for infiltration or attack.
-
Goal of Privacy Protection: Maintain ownership of private information and find countermeasures against threats. In this section, we adopt these terms and discuss privacy concerns in the field of recommender systems (RS).
robustness
While the recommender system improves the efficiency of information search and benefits both the client and the producer, it can also compromise the robustness of the user of the system, which opens the door for third-party injection attacks through configuration files (aka shilling attacks). ) leaves room for manipulating users’ recommendation results. The motives of these attacks are often malicious, such as personal gain of illicit profits, market penetration of certain goods/brands, or even system failure. Since recommender systems are already employed in many high-risk decision-making scenarios, this vulnerability raises concerns about how to safely adopt machine learning techniques in recommender systems, and how to carefully design recommender systems to be robust against attackers [13] ] has robustness and credibility concerns.
controllability
AI controllability is one of the most important problems facing humans [313], it is essential when users interact with intelligent systems, and has been studied in the field of human-computer interaction (HCI) for more than 20 years [11, 276]. In recommender systems that interact with humans [154, 199, 245, 250, 258], the importance of controllability cannot be ignored. However, despite recent successful improvements in recommendation performance, the controllability problem in recommender systems has become a new major problem: most of the current RSs are mostly uncontrollable by system users, who can only passively receive recommendation results . More specifically, when using an uncontrollable recommendation system, users can only passively choose to accept or not to accept the recommendation results, and it is difficult to control what the recommendation results they receive, and more importantly, to control how the recommendation system treats themselves. learn. In fact, controllability is an important aspect of building trustworthy recommender systems. Recent studies have shown that users may be dissatisfied even with high recommendation accuracy [128, 198], and increasing users’ controllability over the recommender system can increase user satisfaction and trust in the recommendation results [133, 142, 146, 153, 197, 305, 340].
in conclusion
This review summarizes the current development and trends of trusted recommender system research, aiming to promote and advance the research and implementation of trustworthy recommender systems in the future . This review provides a roadmap for the comprehensive development of trustworthy recommender systems from a technical perspective. We first define the trustworthiness of recommender systems and characterize them by classifying trustworthiness principles. Subsequently, we introduce and discuss recent advances in trusted recommender systems in terms of interpretability, fairness, privacy, controllability, and robustness. We describe the basic idea of each component, provide a detailed overview of existing methods for each component, and suggest future research directions for these components, especially from a cross-aspect perspective. Overall, the research field of trustworthy recommender systems is important and is thriving with a range of different approaches and applications, and at the same time, making recommender systems responsible and worthy of our trust is something our research community needs to address One of the biggest challenges. We hope that this review will provide researchers interested in this field with sufficient background and knowledge to meet this challenge.
Expert knowledge and convenient viewing
Easy to download , please pay attention to the official account of Zhizhi (click the blue Zhizhi above to follow)
By replying to " TRS " in the background, you can obtain " How reliable is the recommendation system? Rutgers University's latest review of "Trusted Recommender Systems", 43 pages of pdf explaining the composition and technology of trusted RS "Knowledge download link
