A causal intervention framework for synthesizing mobility data and evaluating predictive neural networks

TL;DR

The paper addresses the interpretability and robustness of neural networks for mobility prediction by introducing a causal intervention framework. It uses a structural causal model (SCM) to synthesize interventional mobility traces by perturbing parameters such as $p^{new}$, $n_j$, and $f$, and by manipulating transition dynamics through $ ho$ and $  gamma$. These traces feed next-location predictors (LSTM and MHSA), exposing how sequential transition patterns, exploration propensity, and location preferences drive accuracy and OoD robustness under domain shifts. Open-sourced framework enables benchmarking across environments and datasets, advancing interpretability and robustness of mobility prediction in real-world applications.

Abstract

Deep neural networks are increasingly utilized in mobility prediction tasks, yet their intricate internal workings pose challenges for interpretability, especially in comprehending how various aspects of mobility behavior affect predictions. This study introduces a causal intervention framework to assess the impact of mobility-related factors on neural networks designed for next location prediction -- a task focusing on predicting the immediate next location of an individual. To achieve this, we employ individual mobility models to synthesize location visit sequences and control behavior dynamics by intervening in their data generation process. We evaluate the interventional location sequences using mobility metrics and input them into well-trained networks to analyze performance variations. The results demonstrate the effectiveness in producing location sequences with distinct mobility behaviors, thereby facilitating the simulation of diverse yet realistic spatial and temporal changes. These changes result in performance fluctuations in next location prediction networks, revealing impacts of critical mobility behavior factors, including sequential patterns in location transitions, proclivity for exploring new locations, and preferences in location choices at population and individual levels. The gained insights hold value for the real-world application of mobility prediction networks, and the framework is expected to promote the use of causal inference to enhance the interpretability and robustness of neural networks in mobility applications.

Figures (8)

• Figure 1: Framework for evaluating the robustness of prediction networks through causal interventions. We generate location sequences from mechanistic models and feed them into prediction networks to evaluate the prediction performance (blue arrows). This process is repeated for interventional location sequences, obtained by modifying the distribution of behavioral parameters (green arrows). The differences in mobility patterns and prediction performances are compared to assess intervention strengths and network robustness (red arrows).
• Figure 2: Mechanistic generative model DT-EPR. The individual at location $i$ visited $S=4$ locations with a frequency proportional to the size of the location circle at time $t$. At time $t+\Delta t$, the individual chooses to either explore a new location with probability $p^{new}$, where the next location $j$ will be chosen based on its population attractiveness $n_j$ and the travel distance $r_{i,j}$ (d-EPR mechanism; upper panel), or return to a previously visited location with complementary probability $1-p^{new}$, where the location probability $\Pi_j$ is proportional to the empirical visit frequency from $i$ (IPT mechanism; lower panel). Figure adapted from song_modelling_2010.
• Figure 3: Synthetic location visits generated by EPR-like individual mobility models. (A) Spatial distribution of location visits for an exemplary user. (B) Mobility entropy and (C) motifs proportion distributions of the generated population. Map data ©OpenStreetMap contributors, ©CARTO.
• Figure 4: The entropy and motifs proportion distributions of observational and interventional location sequences. We show the metric distributions for (A) hard interventions on $p^{new}$, (B) interventions on $\rho$ by shifting $\mu\vert_{\rho}$ of $P(\rho)$, and (C) interventions on $\gamma$ by shifting $\mu\vert_{\gamma}$ of $P(\gamma)$.
• Figure 5: Next location prediction performances for interventions on individuals' exploration tendency. We show the variations in Acc$@$1 and MRR for (A) hard interventions on $p^{new}$, (B) interventions on $\rho$ by shifting $\mu\vert_{\rho}$ of $P(\rho)$, and (C) interventions on $\gamma$ by shifting $\mu\vert_{\gamma}$ of $P(\gamma)$.