Uncovering migration systems through spatio-temporal tensor co-clustering

Zack W. Almquist, Tri Duc Nguyen, Mikael Sørensen, Xiao Fu, Nicholas D. Sidiropoulos

Introduction

Human migration is not a collection of isolated, random movements. Instead, it often manifests as "migration systems"—persistent, structured exchanges of people between specific geographic regions over extended periods. Understanding these systems is crucial for urban planning, economic forecasting, and disaster response. However, identifying these systems amidst the noise of annual fluctuations and external shocks has historically been a challenge.

Traditional approaches often rely on Social Network Analysis (SNA), treating migration as a static graph. To capture change, researchers typically analyze a sequence of snapshots. This method frequently fails to maintain consistency; it is difficult to determine if a cluster identified in one decade is the same system in the next, especially when the overall volume of movement changes. Furthermore, many models struggle to account for the directed nature of migration or the fact that a single location can belong to multiple overlapping systems.

Methodology: Spatio-Temporal Tensor Co-Clustering

To address these limitations, this research introduces spatio-temporal (ST) tensor co-clustering. Rather than analyzing a series of 2D matrices, the authors organize migration data into a third-order tensor $\mathcal{X}$ with dimensions representing Origin ($I$), Destination ($I$), and Time ($K$).

The core of the analysis uses Canonical Polyadic Decomposition (CPD), a high-dimensional generalization of Principal Component Analysis. This technique decomposes the massive data cube into a sum of $F$ "rank-one" tensors. Each component represents a distinct migration system characterized by three vectors:

  1. Origin Indicator ($a_f$): Identifies the primary sending locations.
  2. Destination Indicator ($b_f$): Identifies the primary receiving locations.
  3. Temporal Profile ($c_f$): Tracks how the intensity of that specific system evolves over time.
Decomposition of a migration system into origin, destination, and temporal components.

By applying non-negativity constraints, the model ensures that the results are physically interpretable, as migration counts cannot be negative. This separation allows researchers to identify stable spatial patterns even when the total volume of migration fluctuates due to economic cycles or sudden disasters.

Key Insights and Case Studies

The power of this approach lies in its ability to decouple geographic structure from temporal intensity. This leads to several significant findings across different scales.

Detecting Exogenous Shocks: Hurricane Katrina

The method proves exceptionally effective at identifying the impact of "black swan" events. In a case study of Louisiana, the ST tensor model clearly distinguished between two distinct systems: a displacement system (the immediate exodus from New Orleans in 2005) and a recovery system (the subsequent return and arrival of new residents peaking around 2009). While traditional methods merely showed a community "shrinking," the tensor approach revealed that the post-disaster recovery was driven by a different set of counties than the pre-disaster system, indicating a fundamental shift in migration patterns.

The North-South Divide in California

In analyzing California's internal migration, the model identified distinct Northern and Southern systems. Interestingly, it highlighted Los Angeles as a critical "hub" or link between these two systems. Unlike "hard partitioning" methods like $k$-means, which force a city into a single group, the tensor approach allows for overlapping communities. A city like Los Angeles can simultaneously serve as a major sender in one system and a major receiver in another.

On a national level, the model quantitatively captured the "Great American Migration Slowdown"—the steady decline in US internal mobility since the 1990s—within its temporal profiles. On the international stage, it identified six major stable systems, such as the Mexico-US and Eastern European-Russian corridors. It also detected geopolitical shifts, such as the end of the Syrian occupation of Lebanon in 2005, reflected as a significant change in the Middle Eastern temporal profile.

Discussion and Limitations

While the ST tensor co-clustering method offers a robust framework for migration analysis, some challenges remain. The selection of the number of systems ($F$) requires careful tuning; the authors utilized "information decay" to determine $F=6$ for their primary analysis. Additionally, the reliance on IRS tax data introduces potential biases, as it excludes non-filers, such as the very poor or the elderly, and obscures small flows (fewer than 20 households) for privacy reasons.

Future research could integrate these discovered migration systems into formal demographic forecasting models. By understanding the underlying "pulse" of human mobility, planners can better predict future infrastructure and housing needs, moving beyond simple extrapolations of current trends.

Summary updated

BibTeX
@inproceedings{almquist2024uncovering,
  author = {Zack W. Almquist and Tri Duc Nguyen and Mikael Sørensen and Xiao Fu and Nicholas D. Sidiropoulos},
  title = {{Uncovering migration systems through spatio-temporal tensor co-clustering}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-78112-z},
}
Figure