# Large-Scale Estimation and Analysis of Web Users' Mood from Web Search Query and Mobile Sensor Data

## TL;DR

This paper presents a system to estimate web users' moods using search queries and mobile sensor data, revealing mood patterns and their relation to events like the pandemic.

## Contribution

A novel two-step model using search queries and sensor data to estimate mood states at scale, with real-world deployment and mood tracking during the pandemic.

## Key findings

- A nationwide mood score was created, showing mood rhythms and inverse synchronization with new COVID-19 cases.
- The system detected major news events affecting user moods at fine-grained time resolutions.
- Certain ad types were linked to distinct user mood patterns.

## Abstract

The ability to estimate the current mood states of web users has considerable potential for realizing user-centric opportune services in pervasive computing. However, it is difficult to determine the data type used for such estimation and collect the ground truth of such mood states. Therefore, we built a model to estimate the mood states from search-query data in an easy-to-collect and non-invasive manner. Then, we built a model to estimate mood states from mobile sensor data as another estimation model and supplemented its output to the ground-truth label of the model estimated from search queries. This novel two-step model building contributed to boosting the performance of estimating the mood states of web users. Our system was also deployed in the commercial stack, and large-scale data analysis with >11 million users was conducted. We proposed a nationwide mood score, which bundles the mood values of users across the country. It shows the daily and weekly rhythm of people's moods and explains the ups and downs of moods during the COVID-19 pandemic, which is inversely synchronized to the number of new COVID-19 cases. It detects big news that simultaneously affects the mood states of many users, even under fine-grained time resolution, such as the order of hours. In addition, we identified a certain class of advertisements that indicated a clear tendency in the mood of the users who clicked such advertisements.