Fig. 3: Model performance at predicting test flows between municipalities in training states.
From: Human mobility is well described by closed-form gravity-like models learned automatically from data
A–H For each model prediction for two representative states (Florida and Washington; see Fig. S3 for the remaining states), we assess model performance using four different metrics: A, E Common part of commuters; B, F Absolute error; C, G Absolute relative error; D, H Absolute log-ratio. The common part of commuters (CPC) is a global metric. Thus, we have a single value for each metric. For the other three metrics, we show the median, 50% confidence interval (box), and 95% confidence interval (whiskers). Triangles (◂) indicate the best-performing model for each metric (largest CPC or lowest median). See Methods and Text for the definition and discussion of the different metrics. I–L Summary of performance over test flows in the six training states. The performance ratio is defined with respect to the performance of the gravity model with power law decaying dependence on the distance (Gravity pow); values larger than 1 correspond, for all metrics (including CPC), to performance above the Gravity pow model, whereas values smaller than 1 indicate worse performance. Error bars indicate 95% confidence intervals for the means over states. (See Supplementary Table S6 for numerical values for all individual states, as well as summary statistics.) Overall, all models perform similarly in terms of CPC, whereas BMS models and random forests (and deep gravity for absolute error) perform significantly better in the other statistics.
