Fig. 7: Closed-form models with non-population features.
From: Human mobility is well described by closed-form gravity-like models learned automatically from data
A Spearman’s rank correlation among all pairs of features (distance, 19 origin features, and 19 destination features), for all pairs of municipalities in the training states: New York, Massachusetts, California, Florida, Washington, and Texas. B, C Population of the destination versus: B number of food points at the destination; C main road lines at the destination. In both cases, the straight line represents a linear relationship; this is not a fit, and it is provided as a guide to the eye. D–G Performance ratio of models with respect to the gravity power-law model for the training states. As in Fig. 3, the performance ratio is defined so that values larger than 1 correspond, for all metrics (including CPC), to performance above the gravity power-law model, whereas values smaller than 1 indicate worse performance. Error bars indicate 95% confidence intervals for the means over states. The model labeled BMS 39 Feat corresponds to the most plausible model obtained by the BMS when trained with non-population features (Eq. (2)). H–K Same as D–G, but for out-of-sample states. The testing procedure is as in Fig. 6. In general, the BMS 39 Feat model performs slightly better than the most plausible population-only model, but the differences are not statistically significant.
