Recommendations as Challenges: Estimating Required Effort and User Ability for Health Behavior Change Recommendations

Title of the Paper

Recommendations as Challenges: Estimating Required Effort and User Ability for Health Behavior Change Recommendations

Paper Information

• Topic Area: Health recommendation systems, personalized user behavior change
• Keywords: Health recommendation systems, behavior change, user ability, difficulty, personalization, Rasch model, Elo, TrueSkill, Glicko-2

Research Background and Problem

• Problem and Challenges: Can health recommendation systems personalize recommendations based on task difficulty and user ability? Most existing methods (e.g., Rasch model) rely on extensive user survey data and fail to dynamically capture user behavior.
• Importance: Behavior change interventions are crucial for preventing and managing chronic diseases; recommending tasks of appropriate difficulty can enhance user engagement, thereby improving the effectiveness of health behavior interventions.
• Research Motivation and Related Work: The authors reviewed related literature in music, education, and behavior change fields, highlighting a critical research gap in integrating recommendation methods based on user ability and required effort (difficulty).

Solution

• Methods or Solutions:
  1. Proposed five recommendation system methods:
     • Explicitly predefined task difficulty features
     • Three scoring systems from sports and gaming: Elo, Glicko-2, and TrueSkill
     • Rasch model (a difficulty modeling method based on survey data)
  2. Combined long-term online experiments with user interaction data to compare the performance of these methods in estimating user ability and task effort.
• Innovations:
  • First-time application of gamified algorithms (Elo, Glicko-2, TrueSkill) in health recommendation systems to dynamically generate user ability and task difficulty.
  • Introduced a time decay mechanism and combined dynamic user behavior updates for personalized recommendations to enhance user engagement.
• Implementation Steps and Key Technologies:
  • Data Collection: Conducted a two-week experiment to collect empirical user behavior data, forming a user-task interaction matrix.
  • Algorithm Optimization and Comparison: Modeled the five methods and compared their advantages and disadvantages across multiple dimensions, including accuracy, performance, and flexibility.
  • Online Validation: Integrated the best-performing Glicko-2 algorithm into a complete recommendation system framework to test its impact on user engagement and recommendation effectiveness.

Research Outcomes

• Specific Results:
  • Conducted comparative evaluations of five algorithms, identifying Glicko-2 (especially its Hybrid version) as suitable for health recommendation tasks related to behavior change.
  • Developed a health recommendation dataset containing 447 behavior intervention suggestions and explored ability-difficulty modeling using the Rasch model.
  • Designed an application framework compliant with ethical evaluation and data privacy standards, demonstrating the effectiveness of Glicko-2 in dynamically updating user ability and task effort.
• Comparative Advantages Over Existing Solutions:
  • Compared to the Rasch model and TrueSkill, the Glicko-2 algorithm is faster, better suited for dynamic user interaction data, and provides native time decay support.
  • Offers superior dynamic adaptability in matching user ability with recommendation difficulty compared to traditional explicit labeling methods.
• Experimental or Evaluation Results:
  • User experiments showed that recommendation schemes integrating user ability and task effort (based on Glicko-2) significantly increased user engagement (approximately threefold growth).
  • Crucially, users expressed higher satisfaction with highly adaptive personalized recommendations.
• Limitations and Future Directions:
  • Limited Data Scale: The total number of trial users was relatively small (fewer than 200), and interaction data sparsity was high (sparsity level of 0.865).
  • Future work will extend the research scope to other recommendation domains (e.g., education, transportation) and explore the application of methods on larger-scale datasets.
  • Further research is recommended on the dynamic changes and adjustment mechanisms of user ability models in long-term behavior tracking.