Tweether: A Visualization Tool Displaying Correlation of Weather to Tweets

As the generation of social media, we can instantly express how our day is going; however, unknowingly the weather can play a key role in how we are feeling. The weather may dictate our lives regardless of what may be happening. The relationship between weather and mood has been immensely studied to show that the weather does play a major factor regarding our emotions. However, how we visualize the relationship and influence between weather and human emotions remains an interesting question. Based on the natural correlation between weather and mood, we propose Tweether, a real-time weather and tweet visualization tool, to see how Twitter users feel regarding the weather they experience. Our visualization displays a current reflection of emotions in a set of select geographic regions and also predicts possible emotions in these regions in response to the weather forecast. The visualization uses multiple layers to show the connection between geolocations, weather, and emotions. By aggregating multiple users with emotions, we create an aesthetic design in a 3D manner that is relatively free of visual clutter and it is simple to understand the relationships between weather and emotions. Introduction Weather affects our daily lives, from what we wear, what activities we do, what type of transportation we use, what we eat, or even how we feel. With the increasing accuracy of weather forecasts, people can gain an idea on the type of weather they can expect for upcoming days. Activities are usually planned according to the weather outside (e.g., weddings) and alternative plans must be made in case of inclement weather. How people dress is also affected by weather; when the temperature drops people need to wear coats to stay warm. The economy is also greatly affected by the weather. Certain weather conditions can lower crop yield and cause higher prices in stores. Disastrous weather phenomena such as hurricanes, tornadoes, or even floods can cause devastation in communities resulting in homelessness, death, and destruction. Inclement weather can also cause delays in transportation on roads or via flights. We can also choose to ride our bike to work instead of driving the car if the temperature is warm enough. One thing that is an effect of all these items is how we feel: • Are you sad that you cannot enjoy the outdoors due to rain? • Do you love that it’s raining so you can bundle up and read your favorite book? • Do you love the snow because it’s close to Christmas? • Do you hate the winter because you want it to be spring? These feelings are all brought out by the weather outside. One person can feel positive about a certain type of weather and one person can feel negative. Categorizing similar feelings from a large number of people may reveal some useful patterns that can help decision makers or shareholders make more appropriate decisions (e.g., carnival or sport game arrangements) with respect to weather conditions. Social media researchers can also gain benefits from these patterns by getting insights into how population behaviors may change with weather variation. However, rare visualization research exits to show the correlation between human sentiments and weather conditions. Furthermore, it remains a challenging task to depict massive and complex relationships among many objects. In this work, we present a novel tool, named Tweether, a visualization of real-time Twitter and weather data to show the feelings of current users and how their emotions could fluctuate regarding the weather. We also develop a prediction model that predicts the emotions with respect to the weather forecast. Our work showcases a 3D map which highlights select clusters of weather. The correlation of tweets to weather is represented by a graph. We use texture-based edge bundling to visualize the graph for reducing visual clutter. Introducing a clear relationship between weather and tweets, the design presents a natural manner of representing correlation and revealing high-level patterns. By using Tweether, the end-users, such as decision makers and social media researchers, can explore if the classified weather has any correlation to the majority of the population. Tweether can be used to examine if the majority of the population follows certain patterns, and how the overall sentiment changes in regards to weather. In addition, the prediction model predicts the future feelings of given weather so that public preferences could be foreseen for certain decision making. Related Work Visualizations correlating sentiment and weather are highly sparse, and the presence of live visualizations is also non-existent. However, there are studies showing the two portions of this work. Clustering of weather data has been done many times in the past. There are also a vast number of visualizations which indicate the sentiment of different locations. Psychological Studies The natural correlation of weather with emotion has been studied profusely [5, 11, 15, 17]. With various factors among the different research, the conclusions attained were broad. In general, humidity, sunshine, and temperature have the greatest effect on mood [11]. In some research on weather and mood, correlations have been debunked. There is no consistency due to seasons and time spent outside [5]. Having certain emotions regarding the season has strong links to seasonal affective disorder (SAD), where people are depressed in regards to changes of the seasons, which usually occurs during the winter. However, most psychologists believe that the weather has an impact on psychological intentions [15]. When observing serotonin levels in regards to sunshine there were strong relationships to being happy [17]. It has been found that weather may not play a big role in the positive attitude, but the negative attitude can have a correlation with weather [11]. Social Media, Weather, and Emotions Few works present the use of Twitter data for their social media feed and some form of weather data. In comparison to our work, the following research used past data and developed a 2D graph implementation for visualization. Work has been done using two to four years of Twitter data and correlating it with meteorological data from NOAA [8, 18]. Using urban areas in the United States as the area of interest, the tweets are passed to a sentiment analyzer that has a multilevel process. The researchers first determine keywords which are identified from public events (e.g., entertainment or natural disasters), then identify the mood state, and finally assign sentiment scores. To correlate the weather with the tweets they use a Generalized Mixed Model to display the non-linear relationship between emotion and weather. Using multiple variables for weather (temperature, temperature change, precipitation, snow depth, wind speed, solar energy, and hail), they determine the connection to hostility-anger, depression-dejection, fatigue-inertia, and sleepiness-freshness. Their results indicate that the warmer temperatures create an angrier atmosphere, lower depression, and less sleepiness, and they determine the influence of temperature to mood is trivial. Their visualization is limited to graphs [23]. Other than using urban areas in USA the researchers see relationships between temperature, humidity, and atmospheric pressure for tweets in the United States and weather data from Weather Underground. Using Linguistic Inquiry and Word Count for sentiment classification they see a pattern with temperature and emotion of every state in the United States. Using regression analysis, they find that the warmer states have a happier mood than the colder states. Their visualization is limited to a bar chart. These works are limited in visualization and usability study. Using past data is useful for our training and testing model; however, having a live view of what Twitter users feel is what we aim for in this work. Sentiment Analysis in Social Media Sentiment analysis has been studied vastly. There are various methods to detect the sentiment of a sentence. However, in regards to tweets sentences may be incomplete, because a tweet is limited to 140 characters and the need to express oneself is limited to short meaningful phrases. We find abbreviations, neologisms, acronyms, hashtags, emotions, and URL’s throughout most tweets. Certain features need to be extracted and some need to be filtered out. Filtering of URL’s, usernames, and Twitter special words may be needed in certain scenarios [20]. Stop words (i.e., a, an, the, etc.) are also removed due to not adding any extra sentiment information. For classifying the tweets, a number of different methods are used, and the most prevalent one is the Naive Bayes classifier [20, 22]. Emotions are used as basis for sentiment classification for classifying tweets as positive or negative for the training purposes [20, 1]. Visualizations Clustering Although clustering of data has been extensively studied, it remains a non-trivial task to deal with temporal and time series data. Weather data needs to be clustered based on values, proximities, and changes throughout the given time span. There have been various visualization techniques for time series data. Visualizing time series data using spirals for large data sets can better identify periodic structures in data [27]. Using wavelet to transform data along a multi-resolution temporal representation to find clusters with similar trends is a useful method for exploring data in a time series fashion [28]. Applying smooth data histograms for visualizing clusters in self-organizing maps is a simple method for 2D data sets [21]. The mass majority of clustering visualizations uses k-means clustering on the basis of their algorithms [27, 28]. We choose to follow this pattern as well, as k-means is a widely used algorithm that gives reasonably good results [27, 13]. Bundling The correlation between multiple entities (e.g., various weather and