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

TL;DR

Time2Stop is an intelligent, adaptive, and explainable JITAI system that leverages machine learning to identify optimal intervention timings, introduces interventions with transparent AI explanations, and collects user feedback to establish a human-AI loop and adapt the intervention model over time.

Abstract

Despite a rich history of investigating smartphone overuse intervention techniques, AI-based just-in-time adaptive intervention (JITAI) methods for overuse reduction are lacking. We develop Time2Stop, an intelligent, adaptive, and explainable JITAI system that leverages machine learning to identify optimal intervention timings, introduces interventions with transparent AI explanations, and collects user feedback to establish a human-AI loop and adapt the intervention model over time. We conducted an 8-week field experiment (N=71) to evaluate the effectiveness of both the adaptation and explanation aspects of Time2Stop. Our results indicate that our adaptive models significantly outperform the baseline methods on intervention accuracy (>32.8\% relatively) and receptivity (>8.0\%). In addition, incorporating explanations further enhances the effectiveness by 53.8\% and 11.4\% on accuracy and receptivity, respectively. Moreover, Time2Stop significantly reduces overuse, decreasing app visit frequency by 7.0$\sim$8.9\%. Our subjective data also echoed these quantitative measures. Participants preferred the adaptive interventions and rated the system highly on intervention time accuracy, effectiveness, and level of trust. We envision our work can inspire future research on JITAI systems with a human-AI loop to evolve with users.

Figures (9)

• Figure 1: Time2Stop System Overview. The overall interaction flow consists of two loops. The first loop (green) includes: ① The mobile app continuously gathers contextual and app usage data (left) and transmits them to the cloud server. ② On the cloud server's end, feature extraction, ML model inference, and explanation generation occur (right). The ML model output and explanations are sent back to the user. The second loop (loop) includes: ③ In cases where the model predicts "overuse", an intervention would show up while allowing users to provide feedback. The feedback is then forwarded to the cloud server to update the ML model. ④ The updated ML model is subsequently employed to provide more personalized and adaptive interventions.
• Figure 2: In-the-Moment Labeling and Intervention Interfaces. (Left) In-The-Moment Label Collection Interface; (Right) Time2Stop Intervention Interface. It encompasses four key components from top to bottom: (1) typing-based intervention task, (2) ML model explanations highlighting feature categories aligned with the model's output, (3) collection of user feedback -- this is an optional question that users can choose to respond or ignore, and (4) user actions.
• Figure 3: Overview of System Implementation. (Top): Model Inference. (Bottom): Model Update Leveraging User Feedback.
• Figure 4: Field Experiment Flowchart
• Figure 5: Intervention Accuracy (Top) and Receptivity (Bottom) Comparison across Three Intervention Types. Error bar indicates standard deviation. The same below. The two adaptive versions (with and without explanation) are merged into Adaptive to highlight better that adaptive ML-based methods had higher intervention accuracy and receptivity.