Peter Edsberg Møllgaard: How behavioural data such as social media and mobility might help the fight against the pandemic.
COVID-19 has stirred up the everyday life and introduced new problems in our society. It is not the first epidemic/pandemic that we have faced, but advances in technology have created highways for information to spread in real time. Instant access to news and social media informing us about the current situation changes our behavior instantly. Traditional epidemiological methods rely on random spreading through a population without greater changes in human behavior. The observed changes in mobility therefor makes traditional epidemiological methods less effective and calls for new approaches. We wish to tackle this problem by exploring human behavior and develop deep learning algorithms to better understand the underlying patterns and changes in behavior. We aim specifically at two focus areas: mobility and video.
We investigate the changes in mobility during the Corona crisis and try to aid authorities’ decision making by explaining the changes in behavior over time on a deeper level, specifically in Denmark but potentially extending it to other countries as well. On top of that we explore mobility on an individual level by developing deep learning algorithms to predict next locations in movement and create global embeddings of location types, for example, fitness, home, work. In short, an embedding is a transformation of input variables, e.g. words, into a high dimensional vector space. Ideally this transformation should capture meaningful semantics of the input with relative distances, like words who have similar meanings should be close. In our case the algorithm should ideally be able to cluster the same type of locations together in the embedding space unsupervised.
Another aim is to create general video embeddings using a dataset of ~370.000 TikTok videos. In short TikTok is a social media platform where users dance to predefined songs with (often) predefined movements and predefined messages. The goal is to create a general embedding algorithm of videos that works on any type of video from YouTube, TikTok, Vine, or other social media. Optimally the embedding should contain as much information as possible so we would be able to do predictions directly on the embeddings instead of creating another video network. We will then use these video embeddings to explore how social media videos has been affected by Corona and more, depending on the findings in the embedding space.