STUaNet: Understanding uncertainty in spatiotemporal collective human mobility

Zhengyang Zhou, Yang Wang, Xike Xie, Lei Qiao, Yuantao Li

TL;DR

This paper introduces STUaNet, a novel framework for understanding and quantifying uncertainty in collective human mobility predictions. By simultaneously estimating internal data quality and external contextual uncertainties, STUaNet aims to provide more reliable predictions for critical applications like urban event forecasting and crowd management.

What Problem Does This Paper Address?

Predicting human mobility is crucial for many urban applications, from resource allocation to public safety. However, existing methods often struggle to account for the inherent uncertainty in these predictions. This uncertainty can arise from various factors, including data quality issues, complex interactions between contextual factors (like weather or events), and the inherent randomness of human behavior. Traditional approaches to uncertainty quantification in deep learning often rely on statistical estimations or Bayesian methods, which can be computationally expensive and may not effectively capture the spatiotemporal evolution of uncertainty. The paper highlights a critical gap: the lack of methods that can quantify spatiotemporal uncertainty in collective human mobility while also considering the influence of various contexts and the dynamic nature of these uncertainties over time. This is particularly important for applications where extreme events or unexpected fluctuations are of significant interest, as demonstrated by the tragic stampede in Shanghai.

Methods

STUaNet tackles uncertainty quantification through a multi-pronged approach:

  • Spatiotemporal Prediction Module: This module leverages a graph convolutional network (GCN) and a Long Short-Term Memory (LSTM) network to capture both spatial dependencies and temporal patterns in human mobility. A key innovation here is the use of a time-sensitive mobility-involved adjacent matrix that considers geographical proximity and potential mobility transitions.
  • Content-Context Uncertainty Quantification (C2UQ) Module: This is the core of the uncertainty estimation. It comprises three key components:
  • Neural Data Quality Estimation: This module assesses internal data quality by measuring period-wise sequence similarities, identifying data noise and pattern variations.
  • Context Interaction Learning (FM-GCN): This component uses a deep Factorization Machine-GCN to model how various external contextual factors (weather, time, etc.) interact with each other and influence uncertainty. It captures both context interactions and spatial dependencies.
  • Active Hierarchical Uncertainty Learning: To address the lack of explicit uncertainty labels, the paper proposes a "data turbulence" scheme. This involves injecting controllable noise and out-of-distribution (OOD) samples to guide the learning of epistemic and aleatoric uncertainties. Two weak supervised indicators (data quality and spatiotemporal variance) are used to guide this process.
  • Gated Mobility-Uncertainty Re-calibration (GMuR) Bridge: This module adaptively leverages the learned uncertainty to re-calibrate and boost the prediction performance. It learns how uncertainty variations impact predictions, allowing for a more robust final output.

Datasets

The paper evaluates STUaNet on three real-world spatiotemporal human mobility datasets:

  • NYC Taxi: Approximately 7.5 million taxi trip records from January to May 2017.
  • SIP Surveillance: Traffic volume data from 108 regions in Suzhou Industrial Park (SIP) from January to March 2017.
  • California Check-ins (Gowalla): Around 736,000 check-in records from California, clustered into 1,200 regions.

Key Insights and Ideas

  • Uncertainty is Spatiotemporal and Context-Dependent: The paper strongly emphasizes that uncertainty in human mobility isn't static; it evolves over space and time and is heavily influenced by external factors.
  • Active Learning for Uncertainty: The "data turbulence" scheme is a clever way to generate pseudo-labels for uncertainty learning when explicit labels are unavailable, making the learning process more proactive.
  • Decomposing Uncertainty: Distinguishing between internal data quality and external context influences provides a more nuanced understanding of uncertainty sources.
  • Re-calibrating Predictions with Uncertainty: The GMuR bridge highlights that uncertainty isn't just a measure of doubt but can also be a valuable signal to improve the accuracy of the primary prediction task.
  • Task-Related Spatial Misalignment: The authors identify a crucial issue where traditional methods might extract features for prediction and uncertainty separately, leading to misalignments that can reduce effectiveness. STUaNet aims to mitigate this.

Key Results

  • Superior Forecasting: STUaNet consistently outperforms several state-of-the-art spatiotemporal prediction baselines (STG2Seq, STGCN, MDL, DCRNN) across all three datasets, achieving significant improvements in MAPE. For example, it surpasses the best baseline by up to 47.21% on MAPE in the SIP dataset.
  • Effective Uncertainty Quantification: STUaNet achieves the best performance in uncertainty quantification metrics like PICP, significantly improving it on the SIP dataset (from 68.83% to 80.74%).
  • Ablation Studies: The results of ablation studies confirm the importance of each component, particularly the neural data quality estimation and the GMuR bridge, for achieving optimal performance.
  • Qualitative Analysis: Visualizations show that predicted uncertainties often align with prediction errors and highlight areas of high uncertainty during events or adverse weather, demonstrating the model's ability to capture real-world phenomena.

Limitations and Future Work

  • Data Turbulence Sensitivity: While effective, the "data turbulence" scheme's sensitivity to the degree of injected noise and OOD samples might require careful tuning.
  • Computational Cost: Although the paper mentions efficiency compared to ensemble methods, deep spatiotemporal models can still be computationally intensive.
  • Future Work: The authors suggest further exploring the quantified regularities and uncertainties of spatiotemporal data with more theoretical analysis. They also aim to explicitly optimize spatiotemporal predictions by identifying the sources of deductible uncertainties.

Summary updated

Abstract

The high dynamics and heterogeneous interactions in the complicated urban systems have raised the issue of uncertainty quantification in spatiotemporal human mobility, to support critical decision-makings in risk-aware web applications such as urban event prediction where fluctuations are of significant interests. Given the fact that uncertainty quantifies the potential variations around prediction results, traditional learning schemes always lack uncertainty labels, and conventional uncertainty quantification approaches mostly rely upon statistical estimations with Bayesian Neural Networks or ensemble methods. However, they have never involved any spatiotemporal evolution of uncertainties under various contexts, and also have kept suffering from the poor efficiency of statistical uncertainty estimation while training models with multiple times. To provide high-quality uncertainty quantification for spatiotemporal forecasting, we propose an uncertainty learning mechanism to simultaneously estimate internal data quality and quantify external uncertainty regarding various contextual interactions. To address the issue of lacking labels of uncertainty, we propose a hierarchical data turbulence scheme where we can actively inject controllable uncertainty for guidance, and hence provide insights to both uncertainty quantification and weak supervised learning. Finally, we re-calibrate and boost the prediction performance by devising a gated-based bridge to adaptively leverage the learned uncertainty into predictions. Extensive experiments on three real-world spatiotemporal mobility sets have corroborated the superiority of our proposed model in terms of both forecasting and uncertainty quantification.
BibTeX
@inproceedings{zhou2021stuanet,
  author = {Zhengyang Zhou and Yang Wang and Xike Xie and Lei Qiao and Yuantao Li},
  title = {{STUaNet: Understanding uncertainty in spatiotemporal collective human mobility}},
  booktitle = {The Web Conference},
  year = {2021},
  url = {https://openreview.net/forum?id=R8DrJSHXnl},
}
Figure