Twitch is an online live streaming video platform that primarily focuses on video game streaming and esports broadcasts. Twitch users often find that discovering new and interesting channels on Twitch is difficult due to the large amount of content on the site. Users don’t want to sift through thousands of popular streamers just to find the few that they actually like. Although the Twitch platform currently provides recommendations based on channel popularity, they don’t take into account a user’s viewing history, and less popular channels that the user might enjoy often slip through the cracks.
In the fall of 2018, five UC Berkeley students came together to create an application that would provide a more curated Twitch user experience while increasing the visibility of lesser known Twitch streamers. Their solution, Twitchly, leverages historical data and predictive analysis to quickly provide personalized recommendations that users are guaranteed to like and maximize a user’s engagement with new channels. The project involved pulling hundreds of thousands of user records from the Twitch API and training a machine learning model that would provide a list of ten channel recommendations based on Twitch user characteristics, incorporating an even mix of popular streamers and channels with smaller audiences.