🤖 machine learning

Massive-STEPS: Massive Semantic Trajectories for Understanding POI Check-ins -- Dataset and Benchmarks

This paper introduces Massive-STEPS, a large-scale, publicly available benchmark dataset spanning 15 diverse cities with enriched semantic POI metadata and extended temporal coverage, designed to overcome limitations in existing mobility data and facilitate reproducible research in human trajectory modeling.

Original authors: Wilson Wongso, Hao Xue, Flora D. Salim

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wilson Wongso, Hao Xue, Flora D. Salim

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

Every day, billions of people move through the world, stepping from home to work, from a coffee shop to a park, tracing invisible lines across the map. When we track these movements, we are not just counting steps; we are trying to understand the rhythm of human life. This field of study, known as human mobility modeling, looks at where people go and why, using data from the places they visit. These places, called points of interest, are the anchors of our daily routines. By studying the sequences of visits—what researchers call trajectories—scientists can help city planners design better neighborhoods, guide travelers to new spots, and even teach computer programs to act like real people. For years, this work has relied on a specific kind of digital footprint: the check-in. This is a simple digital signal a person sends when they arrive at a location, telling the world, "I am here."

However, for a long time, our understanding of these movements has been built on a shaky foundation. Most of the research in this field has depended on data collected over a decade ago, between 2012 and 2013. It is like trying to understand modern traffic patterns using only maps from ten years ago; many of the buildings have changed, new neighborhoods have emerged, and the way people live has shifted. Furthermore, almost all of this old data comes from just two cities: New York and Tokyo. While these are fascinating places, they do not represent the whole world. A person's daily routine in a bustling Indonesian city or a quiet Australian suburb can look very different from life in Manhattan. To truly understand how people move, we need a picture that is both newer and much broader, capturing the diversity of human behavior across different cultures and continents.

A team of researchers has stepped in to fill this gap with a massive new resource called Massive-STEPS. This is a large collection of real-world check-in data that spans fifteen different cities, ranging from major global hubs like New York and Istanbul to under-explored regions like Jakarta and Kuwait City. Unlike previous efforts that relied on a single snapshot in time, this dataset covers two distinct periods: the original 2012–2013 window and a newer, updated window from 2017–2018. This allows scientists to see how cities have changed over time and how people's habits have evolved. The researchers did not just gather raw numbers; they cleaned the data carefully, removing impossible movements—such as a person appearing in two distant cities within a minute—and added rich details about each location, including its name, address, and specific type. This creates a high-quality map of human movement that is open for anyone to use, ensuring that future discoveries can be verified and built upon by others.

To test how useful this new data is, the team put a wide variety of computer models to work, asking them to predict where a person would go next based on where they had been before. They tested everything from traditional statistical methods to the newest artificial intelligence systems. The results offered a clear picture of what works and what does not. The most successful models were those that could understand the complex relationships between different places, treating the city as a connected web rather than a simple list of stops. Interestingly, the researchers found that some of the newer, more complex artificial intelligence systems struggled when faced with the messy reality of real-world data. These advanced systems often performed worse than older, simpler methods because they were not designed to handle the fact that many people in the dataset only had a few recorded trips. This "cold start" problem, where there is very little history to learn from, proved to be a significant hurdle for the most sophisticated tools.

Perhaps the most surprising discovery came when the team looked at how the nature of a city affects the ability to predict movement. They had expected that cities with more variety would be easier to model, assuming that a rich mix of options would provide clearer patterns. Instead, they found the opposite to be true. Cities where the types of places people visit are spread out evenly—where there is no single dominant category like "shopping" or "dining" that overshadows everything else—were actually much harder to predict. In these diverse environments, human behavior appears less predictable, suggesting that when a city offers a balanced mix of life, the path a person takes becomes more unique and harder to forecast. This insight challenges previous assumptions and suggests that the complexity of a city's layout and the variety of its offerings play a crucial role in how we can understand its residents.

By releasing this dataset and the code used to analyze it, the researchers have provided a new standard for the field. They have moved the conversation beyond a narrow focus on a few wealthy, well-studied cities and opened the door to understanding mobility in a truly global context. The work confirms that while we have made great strides in tracking human movement, the most accurate models must be flexible enough to handle the short, sparse, and varied nature of real-life data. It also highlights that the future of this science lies not just in building smarter algorithms, but in gathering better, more diverse data that reflects the full spectrum of human experience. With Massive-STEPS, the scientific community now has a clearer, more comprehensive lens through which to view the intricate dance of daily life across the globe.

Technical Summary: Massive-STEPS

Problem Statement

The field of Point-of-Interest (POI) trajectory modeling and human mobility research faces three critical limitations that hinder the development of robust, generalizable models:

Temporal Obsolescence and Data Quality: The field is heavily reliant on datasets from 2012–2013 (specifically Foursquare NYC/Tokyo and the Global-scale Check-in Dataset). These datasets suffer from temporal decay, where a significant portion of POIs have closed or changed categories, and the Global-scale dataset contains nearly 50% erroneous entries (e.g., physically impossible travel speeds).
Geographic Bias: Research is disproportionately skewed toward two cities, New York and Tokyo, failing to capture the diversity of human mobility behaviors across different cultural and regional contexts, particularly in low-resource or underrepresented regions.
Reproducibility Issues: Many existing studies lack reproducibility due to undefined geographic boundaries, unavailable datasets, or opaque preprocessing pipelines, making fair comparison and replication difficult.

Methodology

Dataset Construction (Massive-STEPS)

The authors present Massive-STEPS (Massive Semantic Trajectories for Understanding POI Check-ins), a large-scale benchmark dataset derived from the Semantic Trails Dataset (STD). The construction process involves:

Source Data: Utilizing two subsets of STD: the cleaned 2012–2013 data and the newer 2017–2018 data from Foursquare Swarm.
Geographic Scope: Covering 15 diverse cities across East/Southeast/Western Asia, Europe, North/South America, and Australia. This includes under-explored regions such as Jakarta, Kuwait City, and Petaling Jaya.
Semantic Enrichment: Aligning STD POIs with the Foursquare Open Source Places dataset to enrich entries with missing metadata, including POI names, addresses, and geographic coordinates (noting that 13.25% of check-ins lack exact venue-level coordinates and default to administrative region coordinates).
Preprocessing:
- Trajectory Grouping: Check-ins are grouped into trajectories based on an 8-hour time interval ( $\delta\tau$ ).
- City Filtering: Trajectories are filtered using OpenStreetMap GeoJSON boundaries to ensure spatial grounding within specific cities.
- Quality Control: Trajectories with fewer than two check-ins and users with fewer than three trajectories are removed to mitigate sparsity.
- Privacy: Place and Category IDs are ordinal-encoded.
Temporal Span: The dataset covers a total of 24 months across two non-consecutive periods (2012–2013 and 2017–2018), enabling longitudinal analysis.

Benchmark Tasks

The paper establishes a comprehensive benchmark evaluating models across three tasks:

Supervised POI Recommendation: Predicting the next POI(s) a user will visit based on their current trajectory and historical behavior.
Zero-Shot POI Recommendation: Evaluating Large Language Models (LLMs) on their ability to reason about trajectories without task-specific fine-tuning.
Spatiotemporal Classification & Reasoning: Assessing models' ability to classify trajectory patterns (e.g., weekend vs. weekday behavior).

Experimental Setup

Models Evaluated: A wide range of architectures including classical methods (FPMC), recurrent neural networks (RNN, LSTPM, DeepMove, Flashback variants), graph neural networks (GETNext, STHGCN), and LLM-based approaches (GNPR-SID, UniMove, LLM-ZS).
Metrics: Performance is measured using Acc@k (accuracy at top-k) and NDCG@k (ranking quality).
Analysis: The authors investigate the correlation between urban features (e.g., POI category entropy, spatial sparsity, check-in density) and model performance.

Key Results

Model Performance

Supervised Tasks: STHGCN (Spatio-Temporal Hypergraph Convolutional Network) achieved the highest average performance across all 15 cities, followed closely by GETNext. This suggests that graph-based approaches effectively model spatial and temporal dependencies.
Baseline Comparison: Classical recurrent models (RNN, LSTPM) performed moderately, while newer variants like Graph-Flashback and LoTNext did not consistently outperform their predecessors across diverse cities, challenging claims of universal improvement.
Foundation Models: UniMove, a trajectory foundation model, struggled to surpass older baselines, particularly in cities with high cold-start rates (short trajectories), indicating difficulties in generalizing from extremely sparse sequential data.
Zero-Shot Performance: LLMs showed varying capabilities in zero-shot settings, with performance heavily influenced by city-specific mobility patterns and inherent geospatial biases.

Urban Feature Analysis

Category Entropy Hypothesis: The authors propose a new insight contradicting some prior literature: cities with higher POI category entropy (more evenly distributed POI categories) tend to be harder to model. The absence of a dominant POI category makes user behavior less predictable compared to cities with concentrated activity types.
Temporal Shifts: Significant distributional shifts were observed between the 2012–2013 and 2017–2018 periods, with changes in the most visited POI categories (e.g., shifts in Bandung and Kuwait City), validating the need for multi-period datasets.

Significance and Contributions

The paper claims the following contributions to the field:

Addressing Data Gaps: Massive-STEPS provides the first large-scale, reproducible benchmark that spans 15 diverse cities and two distinct time periods, directly addressing the temporal and geographic limitations of existing datasets.
Enhanced Metadata: By integrating Foursquare Open Source Places, the dataset offers rich semantic metadata (names, addresses, coordinates) previously missing in many trajectory datasets, facilitating more sophisticated modeling.
Reproducibility: The authors release the dataset, processing code, and benchmarking scripts publicly, ensuring that geographic boundaries and filtering procedures are transparent and replicable.
New Insights: The benchmark reveals that urban complexity, specifically measured by category entropy, is a significant factor in modeling difficulty, offering a new perspective for future model development.
Facilitating Future Research: By providing a standardized, equitable, and diverse testbed, Massive-STEPS aims to support the development of more generalizable POI models and the advancement of human mobility simulation, including for LLM agents.

The dataset and code are available at: https://github.com/cruiseresearchgroup/Massive-STEPS.