Speaking of the weather: Detection of meteorological influences on sentiment within social media

Weather occurs in everyday discussion, in every language around the globe. Many people believe that weather does impact our mood and behavior. However, to what degree does this hold? We investigate a collection of weather data and link them to social media feeds. Our work focuses on the comparison of three popular sentiment metrics (previous work focused on one metric) connected to commonly used weather measures with precise geographical linking. Our geographical linking method boosts confidence in results compared with previous work. Additionally, our comparison of three sentiment metrics demonstrates how results can easily be misrepresented when using analysis with a single sentiment metric only. We adopt a state-of-the-art Twitter sentiment classifier and a `weather classifier' evaluated on gold standard test sets and live Twitter data. With the evaluated classification models, a stream of over 30 million live tweets is classified and linked to weather data collected from the same period at over 2,000 weather stations. Our analysis indicates a significant link between sentiment measures and several measures of weather, suggesting that human behavior in the domain of social media is impacted by external factors such as weather.