Figure 3.

Correlations between mobility and social behaviour. For each city, we compute the (a) distribution of cosine similarity and (b) predictability using observed edges (coloured lines) and compare with distributions made using randomized edges. We find both mobility similarity and predictability are much higher when using actual social contacts compared with random users. Social similarity is also correlated with mobility similarity. (c) Ranking each user's contacts by number of calls, we find that stronger ties are more geographically similar. (d) Moreover, the more common contacts shared by two users, the more geographically similar those individuals tend to be. Finally, we explore how social behaviour is correlated with mobility. (e) We find that users with more unique contacts tend to visit more unique locations. (f) Users who distribute their calls to contacts more evenly (higher entropy) are more predictable than users with more uneven call distributions. This suggests that users who share social attention more evenly also share locations. Electronic supplementary material, figures S2 and S3 show these results controlling for call frequency.