**Time2Stop: Adaptive and Explainable Human-AI Loop for Smartphone Overuse Intervention** **Authors:** Adiba Orzikulova, Han Xiao, Yukang Yan, Marzeh Ghassemi, Yuntao Wang, Yuanchun Shi, Sung-Ju Lee, Xuhai Xu, Zhipeng Li, Anind K. Dey **Abstract:** This paper introduces Time2Stop, an intelligent, adaptive, and explainable Just-In-Time Adaptive Intervention (JITAII) system designed to reduce smartphone overuse. Time2Stop leverages machine learning to identify optimal intervention timings, introduces interventions with transparent AI explanations, and collects user feedback to establish a human-AI loop, adapting the intervention model over time. An 8-week field experiment (N=71) evaluated the effectiveness of Time2Stop's adaptive and explanation aspects. Results show that adaptive models significantly outperform baseline methods on intervention accuracy and receptivity. Incorporating explanations further enhances effectiveness by 53.8% and 11.4% on accuracy and receptivity, respectively. Time2Stop also reduces app visit frequency by 7.0~8.9%. Participants preferred adaptive interventions and rated the system highly on intervention time accuracy, effectiveness, and trust. **Key Contributions:** - Designed and implemented Time2Stop, an adaptive and explainable JITAII system for smartphone overuse. - Conducted a longitudinal field experiment to demonstrate the effectiveness of adaptive and explainable interventions. - Shared lessons learned and discussed design considerations and ethical concerns in creating AI-based smartphone intervention systems with a human-AI loop. **Background:** - Overview of smartphone overuse intervention techniques and JITAII methods. - Introduction to Explainable AI (XAI) and its role in enhancing intervention transparency and user trust. **System Design:** - Machine Learning for Smartphone Overuse Prediction: Features, label collection, adaptive model updates, and explanation generation. - Intervention Design: Mechanism, timing, user feedback, and model explanations. **System Implementation:** - Context Sensing: Data collection using AWARE platform. - Intervention Interface: Design and implementation. - Machine Learning Pipeline: Model inference, update, and explanation generation. **Field Experiment:** - Experiment Design: Four intervention types (Control, Personalized, Adaptive-wo-Exp, Adaptive-w-Exp) and evaluation metrics. - Procedure: Initial modeling phase, break phase, baseline data collection, and intervention phase. - Participants: 71 participants with high Smartphone Addiction Scale scores. **Results:** - ML Model Comparison: RF model outperformed others with an F1 score of 66.7%. - Intervention Accuracy and Receptivity: Adaptive models significantly outperformed baseline methods. - App Usage Duration and Visit Frequency: Time2Stop reduced app visit frequency by 7.0~8.9%. - User Feedback: Participants preferred adaptive interventions and rated the system highly on**Time2Stop: Adaptive and Explainable Human-AI Loop for Smartphone Overuse Intervention** **Authors:** Adiba Orzikulova, Han Xiao, Yukang Yan, Marzeh Ghassemi, Yuntao Wang, Yuanchun Shi, Sung-Ju Lee, Xuhai Xu, Zhipeng Li, Anind K. Dey **Abstract:** This paper introduces Time2Stop, an intelligent, adaptive, and explainable Just-In-Time Adaptive Intervention (JITAII) system designed to reduce smartphone overuse. Time2Stop leverages machine learning to identify optimal intervention timings, introduces interventions with transparent AI explanations, and collects user feedback to establish a human-AI loop, adapting the intervention model over time. An 8-week field experiment (N=71) evaluated the effectiveness of Time2Stop's adaptive and explanation aspects. Results show that adaptive models significantly outperform baseline methods on intervention accuracy and receptivity. Incorporating explanations further enhances effectiveness by 53.8% and 11.4% on accuracy and receptivity, respectively. Time2Stop also reduces app visit frequency by 7.0~8.9%. Participants preferred adaptive interventions and rated the system highly on intervention time accuracy, effectiveness, and trust. **Key Contributions:** - Designed and implemented Time2Stop, an adaptive and explainable JITAII system for smartphone overuse. - Conducted a longitudinal field experiment to demonstrate the effectiveness of adaptive and explainable interventions. - Shared lessons learned and discussed design considerations and ethical concerns in creating AI-based smartphone intervention systems with a human-AI loop. **Background:** - Overview of smartphone overuse intervention techniques and JITAII methods. - Introduction to Explainable AI (XAI) and its role in enhancing intervention transparency and user trust. **System Design:** - Machine Learning for Smartphone Overuse Prediction: Features, label collection, adaptive model updates, and explanation generation. - Intervention Design: Mechanism, timing, user feedback, and model explanations. **System Implementation:** - Context Sensing: Data collection using AWARE platform. - Intervention Interface: Design and implementation. - Machine Learning Pipeline: Model inference, update, and explanation generation. **Field Experiment:** - Experiment Design: Four intervention types (Control, Personalized, Adaptive-wo-Exp, Adaptive-w-Exp) and evaluation metrics. - Procedure: Initial modeling phase, break phase, baseline data collection, and intervention phase. - Participants: 71 participants with high Smartphone Addiction Scale scores. **Results:** - ML Model Comparison: RF model outperformed others with an F1 score of 66.7%. - Intervention Accuracy and Receptivity: Adaptive models significantly outperformed baseline methods. - App Usage Duration and Visit Frequency: Time2Stop reduced app visit frequency by 7.0~8.9%. - User Feedback: Participants preferred adaptive interventions and rated the system highly on