When Killing-Time Periods Don't Boost Notification Engagement: Insights from a Multimodal Large Language Model Analysis of Smartphone Screen Activity

作     者：Chang, Yung-Ju Cho, Mu-Jung Chi, Yikun Chen, Yu-Chun Lee, Yu-Jen Kao, Kuei-Chun Ram, Nilam Robinson, Thomas Reeves, Byron 

作者机构：Information Systems Northeastern University San JoseCA United States Department of Computer Science University of California Los AngelesCA United States Department of Computer Science National Yang Ming Chiao Tung University Hsinchu Taiwan Research Center for Humanities and Social Sciences Academia Sinica Taipei Taiwan Department of Communication Stanford University StanfordCA United States Department of Psychology Stanford University StanfordCA United States Department of Pediatrics Stanford University StanfordCA United States Department of Medicine Stanford University StanfordCA United States 

出 版 物：《SSRN》 

年 卷 期：2025年

核心收录：

主　　题：Smartphones 

摘      要：Prior research suggests that killing-time periods may be ideal for presenting users with notifications that require attention and engagement. This study examines whether this assumption holds across different types of phone activities. Using a two-week, screenshot-based study with 36 smartphone users, we collected automated screenshots and delivered four categories of notifications—news, advertisements, crowdsourcing tasks, and questionnaires. To obtain users’ screen activity information, we employed a multimodal large language model (MLLM), GPT-4o, to infer five categories of phone activities: communication, media consumption, functional tasks, gameplay, and other activities. The MLLM-generated classifications demonstrated strong reliability, achieving an inter-coder reliability score of 0.877 compared to human coders. Our findings reveal that while killing-time periods generally increase notification engagement, this effect varies by activity type. Specifically, engagement is amplified during communication and gameplay, but diminishes during media consumption and functional tasks. Furthermore, this positive relationship between killing time and notification engagement is more pronounced when screen activity is inferred using broader time frames (e.g., 1 or 5 minutes, rather than 5 seconds). We also observed distinct patterns in users’ screen activities prior to engaged vs. unengaged notifications, with these differences becoming more pronounced during killing-time periods. Finally, users were more likely to respond to task-based notifications than to content-based notifications, suggesting that both timing and notification type play critical roles in driving user engagement. © 2025, The Authors. All rights reserved.