NationalMood: Large-scale Estimation of People's Mood from Web Search Query and Mobile Sensor Data
Tadashi Okoshi, Wataru Sasaki, Hiroshi Kawane, Kota Tsubouchi
TL;DR
The paper tackles the difficulty of estimating real-world user mood in web contexts by proposing a two-stream approach that fuses smartphone sensor data (Sensor Mood Model, SMM) with web search query data (Query Mood Model, QMM). It validates the method on a data-rich pipeline involving 460 participants over 90 days and scales the analysis to about 11 million users, enabling both mood-aware advertising and a Daily National Mood Score. Key contributions include demonstrating mood-dependent advertising signals and tracking national mood rhythms, including responses to the COVID-19 pandemic, within a large-scale web service framework. The work highlights a practical pathway for mood-aware personalization and population-level mood monitoring on large platforms, with potential implications for targeted content and public-health insights.
Abstract
The ability to estimate current affective statuses of web users has considerable potential towards the realization of user-centric opportune services. However, determining the type of data to be used for such estimation as well as collecting the ground truth of such affective statuses are difficult in the real world situation. We propose a novel way of such estimation based on a combinational use of user's web search queries and mobile sensor data. Our large-scale data analysis with about 11,000,000 users and 100 recent advertisement log revealed (1) the existence of certain class of advertisement to which mood-status-based delivery would be significantly effective, (2) that our "National Mood Score" shows the ups and downs of people's moods in COVID-19 pandemic that inversely correlated to the number of patients, as well as the weekly mood rhythm of people.