Fig 1.
Extracting human trajectories from trace data.
Raw data (top left) contains timestamps and geo-coordinates each time each individual is active on the platform (e.g., making a phone call). From these data, the trajectory of the person through space and time can be reconstructed (top right). The bottom figure shows the set of locations (e.g., neighborhoods) in which the individual was observed on each day.
Fig 2.
One individual’s locations over six months.
Each row is a different district, each column is a day; cells are colored black if the individual either made or received a call in that district on that day. The red boxes show the individual’s modal district in each month.
Fig 3.
Step 1 identifies segments where an individual is at a location continuously, with no gaps exceeding ϵ days. Red boxes in the bottom figure are detected segments. Step 2 merges neighboring segments together. Step 3 removes overlap between two segments.
Fig 4.
Trajectories of two migration events detected by our approach.
Each row is a different district, each column is a day; cells are black if the individual makes or receives a phone call from that district on that day. Red boxes indicate location segments. Orange line marks the inferred date of migration. (A) Long-term migration in Rwanda (migrated from Kigali to Nyamagabe on 2008-07-27). (B) International migration (migrated from Canada to United Kingdom on 2014-03-26).
Table 1.
Performance of the six frequency-based algorithms and our proposed segment-based algorithm.
See section “Traditional frequency-based methods” for the details of the six frequency-based methods.
Table 2.
Performance of the two approaches at labelers’ different levels of confidence.
Fig 5.
(A) Distribution of migration date difference between our approach and labelers. (B) Distribution of uncertainty between our approach and labelers. (C) Distribution of destination duration. (D) Distribution of home duration. Our method has high accuracy on the estimated migration dates and home and destination segment length.
Table 3.
Selected tweets of a detected migrant who moved from Virginia to New York on 2014-09-04 based on our approach.
Fig 6.
The effect of parameters on performance.
(A) Maximum gap between consecutive days. (B) Minimum number of days in a segment. (C) Minimum proportion of days in a segment. (D) Migration rate over time in Rwanda. The optimal tuning of these three parameters is ϵ = 7, minDays = 30, and propDays = 0.6 in our sample.
Fig 7.
The effect of propDays on the performance of frequency-based methods.