TLDR: The paper introduces two novel, publicly available datasets for human mobility in Paris and New York City. These datasets combine real GPS trajectories with rich semantic layers, including points of interest, transportation modes, weather, and uniquely, synthetic social media posts generated by Large Language Models. Available in both tabular and RDF formats, they offer a comprehensive resource for advanced research in behavior modeling, urban planning, and AI applications, enabling multimodal and semantic analysis of human movement.
Researchers have introduced two groundbreaking, publicly available datasets that promise to significantly advance our understanding of human mobility. These datasets, covering the bustling cities of Paris and New York, go beyond simple GPS traces by integrating a wealth of contextual and social information, offering an unprecedented resource for AI and urban planning research.
The challenge in human mobility research has long been the scarcity of comprehensive, semantically enriched datasets. Existing resources often lack crucial contextual layers, are outdated, or come with strict privacy and commercial restrictions. To address this, a team of researchers has developed a novel approach, combining real-world GPS data with inferred semantic details and, for the first time, synthetic social media posts generated by Large Language Models (LLMs).
What Makes These Datasets Unique?
At their core, these datasets are built upon real GPS trajectories voluntarily shared through OpenStreetMap. However, their true innovation lies in the layers of enrichment:
- Contextual Layers: Each trajectory is segmented into ‘stops’ and ‘moves’. Stops are associated with nearby Points of Interest (POIs) such as restaurants, parks, or historical sites. Moves are enriched with inferred transportation modes like walking, driving, or taking the subway. Daily weather conditions relevant to the trajectory’s location and time are also included.
- Synthetic Social Media: A pioneering feature is the inclusion of realistic social media posts. These posts are generated by a sophisticated LLM (Llama-3.3-70B-Instruct), carefully prompted to create plausible content based on the characteristics of nearby POIs and synthetic user profiles (including gender, age, ethnicity, and social network). This allows for multimodal analysis, linking movement patterns with simulated social expression.
- Dual Formats: The datasets are available in both traditional tabular formats (Pandas dataframes) and Resource Description Framework (RDF) formats. The RDF representation models entities like users, trajectories, stops, moves, and POIs, along with their relationships, forming a knowledge graph. This enables advanced semantic reasoning and aligns with FAIR (Findable, Accessible, Interoperable, Reusable) data principles.
- Reproducible Pipeline: The entire methodology and tools used to generate these datasets are open-source and publicly available, allowing other researchers to customize and build their own versions.
How Were They Built?
The creation process involved several meticulous steps. First, raw GPS trajectories and Points of Interest were collected from OpenStreetMap, while historical weather data came from Meteostat. This raw data underwent rigorous preprocessing to ensure quality, including de-identifying user IDs, filtering out very short or sparse trajectories, and merging user-specific data.
Next, the trajectories were semantically enriched. This involved segmenting them into stops and moves, associating stops with POIs within a 50-meter radius, and inferring transportation modes for moves using a random forest classifier. Finally, the synthetic social media posts were generated, adding a rich, textual dimension to the movement data.
Also Read:
- HiD2: Generating Realistic and Diverse Traffic Scenarios for Enhanced Autonomous Driving Prediction
- SIMSplat: Crafting Dynamic Driving Scenarios with Natural Language
Impact on Research
These datasets are poised to significantly impact various research domains. They provide a robust foundation for:
- Human Behavior Analysis: Gaining deeper insights into mobility patterns, daily routines, and activity recognition.
- Urban Planning: Informing strategies for traffic management, public transport optimization, and city development.
- Algorithm Development: Serving as benchmarks for evaluating new algorithms in movement prediction, recommender systems, and multimodal data analysis.
- Knowledge Graph Construction: Facilitating the creation of detailed urban knowledge graphs that link spatial, temporal, and social information.
- LLM-based Applications: Enabling the development of advanced AI assistants and chatbots capable of answering complex, natural language questions about mobility data, by combining the structured knowledge graph with LLM capabilities.
By making these semantically enriched datasets and their building pipeline publicly available, the researchers aim to empower the scientific community with a powerful resource for developing more interpretable, explainable, and context-aware approaches to understanding human behavior in urban environments. You can explore the full details of this work in the research paper: Human Mobility Datasets Enriched With Contextual and Social Dimensions.
