Time2Stop: Adaptive and Explainable Human-AI Loop for Smartphone Overuse Intervention

Document Title

Time2Stop: Adaptive and Explainable Human-AI Loop for Smartphone Overuse Intervention

Document Information

• Domain: Human-Computer Interaction, Smartphone Overuse, Explainable Artificial Intelligence (XAI)
• Keywords: Just-in-time adaptive intervention, Smartphone overuse, Explainable AI, Human-in-the-loop, Behavioral intervention, Adaptive learning, Personalization, Human-centered AI, Intervention effectiveness, Mobile health

Research Background and Issues

• Problems and Challenges:

  1. Smartphone overuse has become a serious social issue, affecting individuals' physical health (e.g., headaches, chronic neck pain), mental health (e.g., anxiety and depression), and social well-being (e.g., distraction, family conflicts).
  2. Existing intervention methods (e.g., usage statistics, application access restrictions) are mostly based on simple rules and cannot dynamically adapt to the complex and variable usage behaviors and preferences of individuals.
  3. There is a lack of AI-driven Just-In-Time Adaptive Intervention (JITAI) methods to address smartphone overuse, and the integration of human-AI interaction feedback in AI-driven interventions remains unexplored.
  4. Traditional "black-box" AI models, despite their high predictive capabilities, suffer from poor explainability, leading to a lack of user trust and acceptance.

• Significance:

  1. Intelligent intervention tools are increasingly important for reducing smartphone overuse, especially in the fields of human-computer interaction and mobile health.
  2. Introducing Explainable AI (XAI) into human-computer interaction can enhance user trust and cooperation while improving the effectiveness and transparency of interventions.
  3. Researching how to integrate user feedback to achieve personalized and adaptive interventions can contribute to the development of smarter intervention systems.

• Motivation and Related Work:

  1. Previous studies have made preliminary explorations in the field of self-limitation and intervention for smartphone use, but they are often limited to device-level or application-level approaches and cannot adjust interventions based on dynamic changes in user behavior.
  2. Preliminary research suggests that combining user behavior data with AI algorithms may be effective in selecting dynamic intervention timing, but there is currently a lack of exploration in human-in-the-loop scenarios.

Solution

• Proposed Method:

  1. Developed the Time2Stop system, an AI-based JITAI system that integrates user feedback (human-in-the-loop) to achieve adaptive and explainable system modeling.
  2. The system consists of four main components:
     • A smartphone sensing application for collecting user behavior and contextual data.
     • A cloud-based machine learning pipeline for extracting behavioral features, detecting potential smartphone overuse, and generating explanations.
     • A user intervention interface for delivering interventions upon detecting overuse and collecting user feedback.
     • An ML model iteratively updated using user feedback to achieve long-term adaptability and precise interventions.

• Innovations:

  1. Proposed a human-AI feedback loop system capable of continuously learning and optimizing intervention models as user behavior evolves.
  2. Integrated explainable AI modules to enhance user trust and intervention acceptance by providing high- and low-level explanations for intervention reasons.
  3. Pioneered the combination of dynamic ML model updates and feature explanation functionalities to address smartphone overuse.

• Implementation Steps:

  1. Data Collection: Passively collect user behavior and contextual data through the sensing platform (AWARE), such as app usage time, screen unlock frequency, activity status, and location.
  2. Label Acquisition: Use Ecological Momentary Assessment (EMA) to collect real-time self-reports of smartphone overuse during app usage.
  3. Model Update: Use user feedback as new labels to update the model daily, reflecting changes in user behavior.
  4. Explanation Generation: Apply SHAP methods to compute feature importance and provide simplified, user-friendly explanations of model predictions.

Research Outcomes

• Specific Results:

  1. An 8-week field study with 71 participants demonstrated that Time2Stop significantly outperformed baseline methods in intervention accuracy and user acceptance.
  2. The dynamic adaptive model improved intervention accuracy by 32.8% (relative advantage) and acceptance by 8% compared to traditional methods.
  3. Introducing explanation functionality further enhanced intervention effectiveness, increasing accuracy and acceptance by 53.8% and 11.4%, respectively.
  4. Achieved significant reductions in smartphone overuse, such as a 7.0–8.9% decrease in app access frequency.

• Comparative Advantages Over Existing Solutions:

  1. Adaptive and explainable functionalities significantly improved intervention personalization and transparency.
  2. The integrated solution (Adaptive-w-Exp method) combining dynamic updates and explanation functionalities achieved the highest scores in user preferences, timing accuracy, and trust levels.

• Experimental or Evaluation Results:

  1. Intervention acceptance and effectiveness showed an upward trend during the study period, indicating potential for long-term deployment.
  2. Using more abstract high-level explanations triggered user self-reflection but also caused confusion and distrust among some users.

• Limitations and Future Directions:

  1. Limitations:
     • Participants were primarily young university students, and results may not generalize to other populations.
     • Lack of detailed analysis of specific overuse types (e.g., social media, gaming).
     • Cold-start issues between data collection and intervention deployment remain unresolved, requiring additional calibration time for the system.
  2. Future Directions:
     • Explore more dynamic personalized explanation generation and enhanced human-computer interaction methods.
     • Utilize privacy-preserving technologies such as federated learning to address data privacy concerns.
     • Optimize model update frequency and cold-start mechanisms for seamless real-world deployment.