On the Feasibility of Predicting Users' Privacy Concerns using Contextual Labels and Personal Preferences

Title of the Paper

On the Feasibility of Predicting Users’ Privacy Concerns using Contextual Labels and Personal Preferences

Paper Information

• Domain: Privacy protection and user behavior prediction
• Keywords: privacy prediction, contextual labels, user preferences, privacy attitudes, data practices, contextual integrity, mixed-method research, data privacy
• Conference: CHI Conference on Human Factors in Computing Systems (CHI ’24)

Research Background and Issues

• Issues and Challenges:

  • Users’ privacy concerns are subjective and complex, making them difficult to model accurately.
  • Generic privacy attitude classifications (e.g., Westin’s Privacy Segmentation Index) fail to effectively predict attitudes toward context-specific data practices.
  • Models for transferring user privacy attitudes across different domains remain underdeveloped.

• Significance:

  • Businesses need to understand users’ expectations and concerns regarding their data collection and processing practices.
  • Developers and HCI researchers can leverage predictive models to customize privacy management tools, while policymakers require better insights into public privacy needs to draft relevant regulations.

• Motivation and Related Work:

  • Contextual factors (e.g., purpose of data use, data type) significantly influence users’ attitudes toward privacy. Existing studies are limited to specific environments and struggle to generalize across domains.
  • Therefore, a more universal approach is needed to model user privacy preferences and predict attitudes toward unseen data practices.

Solution

• Proposed Method:

  • ContextLabel: Combines non-exclusive contextual labels with users’ personal privacy preferences to capture the contextual characteristics of privacy data practices at a finer granularity.
  • Develop a label library containing 18 cross-domain, universally applicable labels.
  • Automatically predict users’ potential privacy reactions to other unknown practices based on feedback from a small number of data practices.

• Innovations:

  • Introduces non-exclusive labels to annotate data practices, expanding the contextual information in privacy modeling.
  • Combines free-text feedback with user preferences, reducing reliance on domain-specific experiments.
  • The proposed contextual prediction method achieves high accuracy and can improve prediction performance by updating the label library.

• Implementation Steps:

  1. Conducted a five-day online survey, collecting feedback from 38 participants on 13 data practices.
  2. Annotated data practices with non-exclusive labels, integrating users’ quantitative ratings and qualitative text feedback.
  3. Tested the correlation and effectiveness of labels and user preferences in predicting privacy attitudes.

Research Findings

• Specific Results:

  • The ContextLabel method achieved 73% accuracy in predicting privacy attitudes toward data practices, outperforming the Privacy Segmentation Index (56%) and contextual integrity factors (59%).
  • Identified universally significant labels (e.g., “price discrimination,” “loss of data control”) that are highly relevant to user concern categories.
  • Users’ attitudes toward privacy issues exhibited high consistency over short periods (ICC up to 0.8).

• Advantages:

  • ContextLabel greatly enhances applicability in fine-grained data practice scenarios and cross-scenario prediction capabilities compared to existing methods.
  • The method supports easy label updates, improving the scalability of research outcomes.

• Experimental or Evaluation Results:

  • Users demonstrated the ability to evaluate privacy attitudes logically.
  • Non-exclusive labels provided more intuitive and accurate privacy information modeling than contextual factors and traditional privacy segmentation indices.

• Limitations and Future Directions:

  • The current label set is based on preliminary testing and requires additional detailed classifications to capture more privacy scenarios.
  • The method’s performance over long-term spans or among more diverse demographic groups has yet to be validated.
  • Future research could expand study contexts, diversify experimental designs to enhance generalizability, and further explore the method’s application in addressing the privacy paradox.

Remarks

This paper’s method emphasizes capturing fine-grained details and improving modeling accuracy, offering valuable insights for the design and implementation of future privacy protection technologies.