🔬 physics

Predictability of Human Movements across Industry Sectors using Multilayer Networks

This study demonstrates that nonlinear machine learning models, particularly random forest regression, effectively predict human movement patterns across diverse industry sectors by leveraging demographic and infrastructure features, with inward movements generally being more challenging to forecast than outward ones.

Original authors: Maisha Islam Sejunti, Melissa Butler, Yingjie Hu, Dane Taylor

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maisha Islam Sejunti, Melissa Butler, Yingjie Hu, Dane Taylor

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the next move of a giant, invisible crowd. This isn't about guessing where a single person might go for lunch; it's about understanding the massive, flowing rivers of people that move through our cities every day. Scientists call this "human mobility," and it's like studying the heartbeat of a city. If we can understand these patterns, we can design better cities, stop diseases from spreading like wildfire, and even plan how to help people during emergencies. To do this, researchers often use "networks," which are like maps where dots represent neighborhoods and lines represent the trips people take between them. But here's the twist: people don't just move randomly. They move to specific places for specific reasons—some go to schools, others to hospitals, and some to grocery stores. By separating these trips into different "layers" based on where people are going (like peeling back layers of an onion), scientists can see a much clearer picture of how our society actually works.

In this study, a team of researchers decided to play a game of "guess the crowd" using a slice of Texas called Harris County. They wanted to see if they could use computers to predict how many people would move into or out of different neighborhoods, depending on whether those people were heading to a school, a restaurant, a hospital, or a store. They built a digital model using a "multilayer network," which is like a stack of transparent maps, each showing a different type of trip. They fed this model a bunch of clues, like how many people live in an area, how much money they make, and how many buildings of a certain type (like schools or shops) exist there. Then, they pitted ten different computer programs against each other to see which one was the best at guessing the numbers.

The results were a bit like a surprise party. The researchers found that the "smartest" computers—specifically the ones that use complex, non-linear thinking called "Random Forest Regression"—were much better at guessing than the simpler, straight-line math models. It turns out that predicting where people are coming from (outgoing trips) is relatively easy; if you know how many people live in a neighborhood, you can usually guess how many will leave. But predicting where people are going to (incoming trips) is much harder, like trying to guess which store a stranger will walk into. The only time the "incoming" prediction got easy was when the destination was a restaurant or a grocery store; people seem to have a very predictable hunger!

The team also discovered that the type of destination matters a lot. Movements to places like schools and "other services" were the hardest to predict, likely because they depend on tricky things like school holidays or sudden changes in plans. In contrast, trips to food and retail spots were the easiest to forecast. A key finding was that for the weekly, moving targets, the smart computers needed to know what time of year it was. When the researchers gave the models a "calendar" (using math to represent the 52 weeks of the year), the predictions got a huge boost, especially for schools, which have very strong seasonal rhythms. Without this time-traveling clue, the models struggled to guess the drops in traffic during summer or winter breaks.

Ultimately, the study suggests that while we can't perfectly predict every single human step, we can get very good at guessing the big picture, especially if we use the right kind of computer brain and pay attention to the season. This isn't just a game; it's a practical step toward helping city planners, doctors, and emergency responders make smarter decisions. By understanding which industries draw crowds and when, we can build better transportation systems, prepare for disease outbreaks, and make sure our cities work for everyone. The paper doesn't claim to have solved the mystery of human movement forever, but it does show us a powerful new way to peek behind the curtain of our daily lives.

Technical Summary: Predictability of Human Movements across Industry Sectors using Multilayer Networks

Problem Statement
Understanding the spatiotemporal patterns of human movement is critical for diverse applications ranging from urban design and disease control to emergency response. While recent work has integrated mobility data into machine learning (ML) to predict urban dynamics, existing studies have primarily focused on spatial-temporal data for individual future locations or aggregate inflow/outflow. There is a notable gap in predicting node-level structural properties (specifically in-degree and out-degree) within human mobility networks, particularly when these movements are stratified by destination industry sectors (e.g., schools, hospitals, restaurants). Furthermore, existing ML/AI studies for movement prediction insufficiently address the variability across different categories of human movement.

Methodology
The authors construct time-varying, multilayer mobility networks to model human movements in Harris County, Texas, over a 103-week period (January 2018 – December 2019).

Data Sources:
- Mobility & Infrastructure: The SafeGraph dataset provides weekly GPS-derived movement patterns from home census block groups to destination points of interest (POIs). Destinations are classified using the 2017 North American Industry Classification System (NAICS). The study focuses on eight specific industry sectors (e.g., Retail Trade, Health Care, Education, Food Services) and a layer-aggregated network representing all movements.
- Demographics: U.S. Census data (2019 American Community Survey) provides six socioeconomic features: population, population density, income, unemployment rate, percentage of non-white population, and poverty rate.
Network Construction: Nodes represent census tracts. Edges in different layers represent the weekly count of movements from one tract to another where the destination belongs to a specific NAICS industry code. The study analyzes both time-averaged node degrees and weekly varying node degrees.
Predictive Modeling:
- Task: Predicting the in-degree (movements into a tract) and out-degree (movements out of a tract) for each census tract, stratified by industry sector.
- Models: Ten statistical and machine learning models were compared: four linear models (Multivariate Linear Regression, ElasticNet, Ridge, Huber) and six nonlinear models (Random Forest Regression, XGBoost, Extra Trees, Nearest-Neighbor, Gradient Boosting, Multilayer Perceptron).
- Features: Input variables included infrastructure counts (number of buildings per industry type in the destination tract) and demographic/socioeconomic factors. For weekly predictions, two time-encoding features (sine and cosine of the week number) were added to capture seasonal cycles.
- Evaluation: Model performance was measured primarily using the coefficient of determination ( $R^2$ ), supplemented by RMSE, MAE, and MAPE. Feature importance was quantified via regression coefficients (for linear models) and impurity decrease (for Random Forest).

Key Results

Model Performance: Nonlinear models consistently outperformed linear models. Random Forest Regression (RFR) was identified as a consistent top performer across most tasks. For weekly predictions, nonlinear models achieved up to a 36% increase in accuracy for out-degree prediction and 15% for in-degree prediction when time-encoding variables were included, compared to time-averaged values.
In-Degree vs. Out-Degree: Predicting in-degrees (inward movements) was generally more difficult than predicting out-degrees (outward movements) across all industry sectors and model types.
Feature Importance:
- Out-Degree: Population size was the most influential predictor for outward movements across all sectors.
- In-Degree: Infrastructure counts specific to the destination industry were the most critical predictors for inward movements. For example, predicting movements to hospitals relied heavily on the count of healthcare infrastructure in the destination tract.
- Cross-Sector Influence: Infrastructure for food services and retail trade emerged as important predictors even for movements to other industry types.
Sector-Specific Predictability:
- Easiest to Predict: Movements associated with food services and accommodation (NAICS 72) and retail trade (NAICS 44) were the most predictable.
- Hardest to Predict: Movements to education services (NAICS 61) and "other services" (NAICS 81) were the most difficult to predict.
- Seasonality: The inclusion of time-encoding variables significantly improved predictions for sectors with strong seasonal dynamics, particularly education. The omission of these variables caused the largest drop in accuracy for education-related movements, attributed to the complex nonlinear patterns of school breaks.

Significance and Claims
The authors claim this work provides a practical step toward using machine learning for human movement modeling by explicitly addressing the predictability of movements across different industry sectors. The study demonstrates that:

Nonlinear models are superior for capturing the complex, heterogeneous nature of human mobility, especially when temporal dynamics are involved.
Predictability is asymmetric: Outward movements are more systematically driven by aggregate characteristics (like population), while inward movements depend on complex, destination-specific factors (like infrastructure) and are inherently harder to model.
Sector specificity matters: Different industry sectors exhibit distinct levels of predictability, with food/retail being highly predictable and education being highly variable due to seasonal factors.

These findings offer practical insights for downstream applications such as urban planning, transportation system design, and emergency response, suggesting that organizations can better anticipate demand fluctuations and allocate resources by understanding the specific predictability characteristics of different industry sectors. The authors modestly note that while their study focuses on a stable period in Harris County, the methodology provides a framework for extending these analyses to other regions and spatial scales.

More like this