The rising safety concerns of deep recommender systems

Recommender systems (RSs) have become essential tools for managing the exponential growth of information available online. By analyzing user behavior patterns and inferring user preferences, these systems significantly enhance user experience by providing personalized recommendations. RSs are applied across diverse domains, from e-commerce platforms to content streaming services, achieving remarkable improvements in user engagement, customer satisfaction, and business performance.

Since the release of collaborative filtering-based recommender algorithms, RSs have flourished over the past 30 years, as illustrated in Figure 1A. With the development of internet technologies, recommender algorithms have evolved from rule-based approaches to those based on shallow models like matrix factorization and have now entered the “deep RS” era, dominated by deep neural network models. The ever-increasing scale and complexity of data and models lead to significant improvements in recommendation performance. However, deep RSs also raise rising safety concerns, including filter bubbles, robustness, and fairness.

As illustrated in Figure 1B, deep RSs collect training data from user interactions and train models by fitting the users' behavioral preferences in the data, finally providing recommendation results to users. During the collecting phase, due to the inherent openness of RSs, attackers can easily inject fake users into the training data, aiming to manipulate recommendations. For instance, Facebook employees found more than 30 cases of politcal manipulation involving 25 countries on the platform, distorting the distribution of item exposure and degrade the overall quality of the RS. Such attacks can distort the distribution of item exposure and degrade the overall quality of the RS.1 Additionally, blindly fitting data during the training process may introduce discrimination into recommendation algorithms, leading to inaccurate or lower-quality recommendations for disadvantaged groups. For example, a study on China’s four largest job recommendation platforms revealed that the algorithms exhibited gender bias, tending to recommend lower-paying positions to female job seekers.2 During the serving stage, the recommendation results ultimately influence user interactions. Over time, this could potentially trap users in filter bubbles, i.e., content that reinforces users’ existing interests, leading to narrow worldviews or even polarization. For instance, an audit of YouTube revealed that following the recommendation trail increases the likelihood of certain ideological users continuing to watch ideologically aligned videos.3 These issues severely affect user experience and the long-term development of deep RSs, warranting further in-depth exploration by researchers.