When Recommender Systems Snoop into Social Media, Users Trust them Less for Health Advice

Title of the Paper

When Recommender Systems Snoop into Social Media, Users Trust them Less for Health Advice

Paper Information

• Subject Area: Artificial Intelligence Recommender Systems, User Privacy, and Trust Psychology Research
• Keywords: Recommender Systems, Health Advice, Social Media, User Trust, Privacy Protection, Personalization, User Choice, Identity Threat

Research Background and Problem

• Problems or Challenges Identified by the Authors:

  • Recommender systems are widely applied in the health domain, but using social media data for personalized recommendations may lead to a decline in user trust.
  • The use of social media data may result in identity threats and raise privacy concerns.
  • Users generally trust health recommendations based on their explicit personal preferences more than those based on data from others or friends.

• Why This Problem is Important:

  • Health recommender systems aim to promote behavioral change, but reduced user trust in the system may directly impact its effectiveness.
  • While integrating social media data is technically feasible, it may have negative psychological effects on users, hindering the widespread adoption of recommender systems.

• Research Motivation and Related Work:

  • Existing studies indicate that highly personalized recommendations can effectively promote health behavior change, but the actual effectiveness of social media-based recommendations in the health domain remains unclear.
  • Related research suggests that user acceptance of health advice may be influenced by privacy and identity-related issues triggered by recommendation algorithms.

Solution

• Proposed Solution or Method by the Authors:

  • Design and evaluate six different personalization methods and experiment with the effects of allowing users to choose their preferred recommendation method.
  • Conduct experiments to test the differences in user trust, identity threat, and privacy concerns between recommender systems based on users' own preferences and those based on social media data.
  • Use the Human-Agent Interaction Influence-Time (HAII-TIME) model to analyze how system prompts and user interactions affect trust and experience.

• Innovative Aspects of the Solution:

  • Emphasizes the positive role of user choice in mitigating identity threats and enhancing trust.
  • Proposes design principles for protecting user identity and privacy, providing guidance for the design of health recommender systems.
  • Systematically investigates the comprehensive psychological impact of recommendation algorithms on users, including identity threats, privacy concerns, and trust.

• Implementation Steps:

  • Employ experimental design, randomly assigning 341 participants to 12 experimental conditions.
  • Collect data on user interactions with the recommender system and analyze their psychological responses and behavioral intentions.
  • Use surveys to quantify user experiences, including trust, identity threat, and privacy concerns.

Research Findings

• Specific Findings:

  • Users exhibited the lowest trust in recommender systems based on social media friends' activities, accompanied by high levels of identity threat and privacy concerns.
  • Recommender systems based on users' own preferences elicited the lowest identity threats, thereby increasing user trust in the system.
  • Providing users with the ability to choose their personalization method significantly enhanced their sense of autonomy and reduced identity threats.

• Advantages Compared to Existing Solutions:

  • Clearly identifies potential issues with social media-based recommendation methods in the health domain and provides empirical support for designing more effective recommender systems.
  • Explores user psychological mechanisms in depth, offering scientific evidence for designing user-friendly health recommender systems.

• Experimental or Evaluation Results:

  • The feature allowing users to choose did not significantly impact system usability, and users' use of the choice feature did not noticeably increase interface complexity.
  • The impact of personalization methods on privacy risks and trust was validated through structural equation modeling.

• Limitations and Future Directions:

  • The experimental recommendation content was uniform, which may not fully reflect the impact of recommendation quality on user evaluations in real-world settings.
  • The study focused solely on preventive health advice scenarios; future research should extend to other recommender system domains to validate the external validity of the findings.
  • The sample size was relatively small, and some conclusions may lack sufficient statistical power, necessitating further expansion of the sample size.

Conclusion

This study, through experimental design, reveals the potential negative psychological impacts on users when recommender systems integrate social media data. It also highlights the importance of user choice in improving user experience and enhancing trust. The findings provide valuable insights for the design of health recommender systems. Future research should focus on cross-domain applications and real-world experiments to further optimize recommendation methods and user experience.