Sentiment Analysis of Twitch.tv Livestream Messages using Machine Learning Methods

Livestreaming refers to the practice of broadcasting media over the internet in real-time. It has steadily grown in popularity with platforms such as Twitch.tv seeing a consistent influx of users over the past few years. Twitch provides a chat feature that allows viewers to participate by sharing their opinions in real-time. These messages can serve as rich and interesting banks of public thought that can help streamers make decisions synchronously about the content that they are producing. However, with several thousand viewers involved, the added complexity of context-specific emotes, memes and the meta-language that pervades through these platforms, summarizing and extracting useful information from them becomes a daunting task. Sentiment analysis is a possible solution to this problem. In this paper, a methodology to perform sentiment analysis with a set of machine learning-based models on livestream messages from Twitch.tv is proposed. Machine Learning models like Support Vector Classifier, Logistic Regression, Decision Tree Classifier, Random Forest Classifier and Multinomial Naïve Bayes were implemented. Among these models, the Support Vector Classifier outperformed the current state-of-the-art model and displayed a 10.3% rise in accuracy, a 9.1% rise in recall and a 7.5% rise in the F1-score.