Mixing Individual and Collective Behaviours to Predict Out-of-Routine Mobility

Sebastiano Bontorin, Simone Centellegher, Riccardo Gallotti, Luca Pappalardo, Bruno Lepri, Massimiliano Luca

Abstract

Predicting human mobility is straightforward when individuals follow established routines, but performance drops significantly during "out-of-routine" events. This paper introduces a hybrid model that dynamically blends individual historical patterns with collective behavior data. By using Shannon entropy to weight these components, the model effectively predicts novel transitions where traditional individual-based models and deep learning architectures often fail.

Introduction

Most human mobility models excel at predicting routine behavior, such as a daily commute from home to work. However, these models struggle with "out-of-routine" mobility—moments when individuals explore new neighborhoods, visit different cities, or when global events like the COVID-19 pandemic disrupt daily schedules.

The fundamental challenge is data sparsity: individual trajectories are often too sparse to predict novel transitions. When a person visits a location for the first time, an individual-based model has no historical data to draw from. This research posits that while an individual's choice might be "new" to them, it likely follows broader collective trends that can be captured by aggregating the behavior of millions of others.

Comparison of sparse individual origin-destination (OD) flows (A) with dense collective flows (B) in Boston.

Methodology

The authors propose a hybrid model, $M$, which integrates two components:

  1. Individual Model ($I$): A Markov-based predictor trained on a specific user's history.
  2. Collective Model ($C$): A Markov-based predictor trained on the aggregated transitions of all users in a city.

The core innovation is a dynamic weighting mechanism based on the normalized Shannon entropy ($S$) of a user’s current location. If a user's history at a location is highly predictable (low entropy), the model trusts the individual pattern. If the history is erratic or non-existent (high entropy), the model shifts its weight toward collective behavior. The final prediction is a linear combination: $M = (1 - S)I + SC$.

The hybrid model integrates individual history and collective patterns using a dynamic weighting mechanism.

Data and Context

The study utilizes a massive, privacy-enhanced GPS dataset from Cuebiq, covering millions of anonymized users across New York City, Boston, and Seattle from January to August 2020. This timeframe is particularly valuable as it includes the pre-pandemic period and the massive mobility shifts caused by COVID-19 lockdowns. Cities were tessellated into GeoHash level 6 tiles (approx. 1.2 km x 0.6 km), and Points of Interest (POIs) from OpenStreetMap were used to correlate mobility with urban features.

Results and Key Insights

The hybrid model $M$ achieved a Top-5 Accuracy (ACC@5) of approximately 0.68 to 0.75 across the three cities, outperforming both purely individual and purely collective models.

Key findings include:

  • The Crowd as a Safety Net: Collective information is the primary driver for predicting unconventional choices. When individuals break routine, they often gravitate toward "collective attractors" like downtown hubs or transit centers.
  • Out-of-Routine Robustness: In scenarios with very low trajectory overlap (0-20%), the hybrid model showed relative improvements of +12% to +15% over individual models.
  • Generalization vs. Memorization: While Recurrent Neural Networks (RNNs) are powerful, they tend to memorize training data. The hybrid model $M$ significantly outperformed deep learning baselines on novel, out-of-routine transitions.
  • Crisis Resilience: During the COVID-19 pandemic, individual routines vanished, causing purely individual models to lose nearly half their predictive power. The hybrid model remained robust by leveraging the collective urban constraints that still governed movement.

Discussion and Limitations

While effective, the model's performance is tied to the density of POIs; in rural areas with weaker collective signals, gains may be less pronounced. Additionally, the model currently treats all out-of-routine trips identically, failing to distinguish between permanent life changes (e.g., a new job) and one-off explorations. Future research could integrate these entropy-based mechanisms into deep learning frameworks to combine high-dimensional feature extraction with robust generalization.

Summary updated

Abstract

Predicting human displacements is crucial for addressing various societal challenges, including urban design, traffic congestion, epidemic management, and migration dynamics. While predictive models like deep learning and Markov models offer insights into individual mobility, they often struggle with out-of-routine behaviors. Our study introduces an approach that dynamically integrates individual and collective mobility behaviors, leveraging collective intelligence to enhance prediction accuracy. Evaluating the model on millions of privacy-preserving trajectories across five US cities, we demonstrate its superior performance in predicting out-of-routine mobility, surpassing even advanced deep learning methods. The spatial analysis highlights the model's effectiveness near urban areas with a high density of points of interest, where collective behaviors strongly influence mobility. During disruptive events like the COVID-19 pandemic, our model retains predictive capabilities, unlike individual-based models. By bridging the gap between individual and collective behaviors, our approach offers transparent and accurate predictions, which are crucial for addressing contemporary mobility challenges.
BibTeX
@inproceedings{bontorin2025mixing,
  author = {Sebastiano Bontorin and Simone Centellegher and Riccardo Gallotti and Luca Pappalardo and Bruno Lepri and Massimiliano Luca},
  title = {{Mixing Individual and Collective Behaviours to Predict Out-of-Routine Mobility}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2025},
  doi = {10.1073/pnas.2414848122},
}
Figure