Figure 3

We present the geolocated Twitter timelines of 6 representative Twitter users. We observe that tweet behaviour is bursty, and that users take many short trips to other locations. Users A–D were classified by both algorithms as migrants, however only users A and B were confirmed by our annotators to display an actual migration event. User A mentions directly moving to Amsterdam from Great Britain, while travelling for business. At one point they mention (paraphrased), “maybe one day I’ll be at my new home for more than 2 days!”. This user is a migrant, but seemingly geo-located in both the Netherlands and Britain over an extended period. User B was identified as a student migrant. While both algorithms identified several moves, only the initial move from the Netherlands to Mexico was identified by our annotators as a true migration event. In both cases, it is possible that a user was using a VPN to access restricted content from another country. User C is a resident of Canada originally from Mexico. Despite the geolocation of the tweets, there was no mention of the user actually being in Mexico, or moving back to Canada. User D likely moved from Germany to the United States, however due to the overlap in time periods, neither algorithm identified the correct move window: during the time window flagged as the time of move the user had not yet made the move to the United States, but had only returned for a brief family visit. The algorithm was, however, correct in principle. Finally, users E and F were not identified as migrants by either of the algorithms. User E is an artist who often travels to Israel to paint, but mentions living in Montreal, Canada. User F is a business professional who takes occasional trips to the United States for work