Just Do Something: Comparing Self-proposed and Machine-recommended Stress Interventions among Online Workers with Home Sweet Office

Title of the Paper

Just Do Something: Comparing Self-proposed and Machine-recommended Stress Interventions among Online Workers with Home Sweet Office

Paper Information

• Research Area: Human-Computer Interaction (HCI), Stress Management, Online Health Interventions
• Keywords: Stress, Intervention, Online Workers, Machine Learning, Self-proposed Solutions, Self-management, Multi-Armed Bandit Algorithm, Micro-interventions

Research Background and Problem Statement

• Problems and Challenges:
  • Work-related stress directly impacts workers' productivity, job satisfaction, personal life, and mental and physical health.
  • Online workers often lack professional support and resources to manage work-related stress.
  • Personalized stress intervention methods lack adaptability and flexibility, especially in diverse stress-triggering scenarios.
• Research Significance:
  • This study aims to explore how browser-based applications can provide micro-interventions to alleviate daily stress for online workers.
  • It focuses on comparing the effectiveness of machine-recommended interventions and user-proposed interventions.
• Motivation and Related Work:
  • Machine learning algorithms have been successfully applied in health interventions (e.g., physical activity recommendations).
  • However, there is a lack of detailed research on whether machine-recommended interventions outperform user-proposed solutions.
  • Traditional stress management applications (e.g., mobile apps like PopTherapy) suffer from high user attrition rates, necessitating improvements in recommendation strategies and user experience design.

Proposed Solution

• Methods and Solutions:
  • Developed a browser extension—Home Sweet Office (HSO)—to provide theoretically supported stress management micro-interventions.
  • Intervention content spans four psychological therapy techniques: cognitive-behavioral, positive psychology, metacognitive, and somatic interventions.
  • Designed and tested a Multi-Armed Bandit (MAB) algorithm to recommend interventions, dynamically adjusting recommendations based on user characteristics and past effectiveness.
• Innovations:
  1. First browser-based stress intervention system, avoiding the excessive attention demands of mobile applications.
  2. Compared the effectiveness of self-proposed, random, and machine-recommended interventions.
  3. Integrated user-preferred custom interventions with algorithmic optimization to enhance personalization.
• Implementation Steps:
  • Developed and deployed the HSO browser extension, allowing users to complete self-assessed stress interventions via the plugin.
  • Designed an intervention pool containing 160 items, with content continuously updated based on user feedback on stress levels.
  • Experimental design included four groups: control group (no intervention), HSO-Self (self-proposed), HSO-Random (randomly selected interventions), and HSO-Bandit (machine-recommended interventions).

Research Findings

• Key Results:
  1. No significant long-term (multi-week) stress improvement was observed in any intervention group, but immediate stress relief was significantly better in the HSO-Self and HSO-Bandit groups compared to the HSO-Random group.
  2. Interventions recommended by the HSO-Bandit group were more diverse and engaging, with longer completion times positively correlated with stress relief.
  3. HSO-Self group users predominantly chose somatic interventions (e.g., breathing exercises, dietary adjustments).
• Comparison with Existing Solutions:
  • Machine-recommended interventions were as effective as user-proposed interventions for immediate stress relief and outperformed random interventions.
  • Unlike mobile apps like PopTherapy, which do not integrate user-proposed interventions, HSO significantly improved intervention diversity and completion efficiency.
• Experimental Data and Evaluation Results:
  • The experiment involved 58 participants, with 1,028 effective interventions completed.
  • Analysis showed that stress relief was more effective when participants voluntarily initiated intervention tasks rather than being prompted by the system.
  • User preference analysis revealed that somatic interventions (e.g., deep breathing or meditation) were generally the most effective.
• Limitations and Future Directions:
  • Due to random assignment technical issues, some users were exposed to default HSO-Self interventions early in the experiment, potentially affecting participation in subsequent groups.
  • The interventions did not directly address specific stress sources (e.g., financial stress, interpersonal issues), which future research could explore.
  • The Multi-Armed Bandit algorithm in the application could be further optimized to avoid user attrition caused by repetitive recommendations.
  • Future studies should extend the research period to evaluate long-term stress changes and explore applicability across different types of online workers.

This study provides valuable insights and implementation strategies for future personalized stress intervention designs by comparing the immediate and long-term effectiveness of user-proposed and machine-recommended interventions in stress management.