RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

RTSKG is a new rail transit station knowledge graph dataset that explicitly models the spatial and semantic interactions among different kinds of urban entities to support city-level rail transit station related tasks. The dataset integrates heterogeneous urban entities—such as rail transit stations, road segments, points of interest (POIs), administrative boroughs, functional areas, and urban blocks—under a specially designed unified ontology. RTSKG consists of two sub-KGs for New York City and Chicago, built from rail transit data, administrative division data, road data, and POI data. The ontology defines ten classes (e.g., Borough, Functional Area, Station, Station Area, Line, Road, Block, POI, POI Category, Road Category) and nineteen relations categorized into geographic location, geographic adjacency, geographic intersection, and non-geographic relations. Station areas are constructed using isochrone regions with 5-minute and 10-minute walking times from station entrances/exits. The dataset contains 238,839 instances and 1,113,638 relation triples for New York City, and 151,845 instances and 448,716 relation triples for Chicago. The authors evaluate RTSKG on two tasks: station-area store recommendation and knowledge-enhanced ridership prediction. For knowledge graph embedding, they compare eleven KGE models and find that GIE performs best on link prediction, likely because the ontology contains both hierarchical and cyclic structures. For station-area store recommendation, RTSKG outperforms existing urban KGs (UUKG and HUSK) across all metrics. For traditional knowledge-enhanced ridership prediction, enhancing models with RTSKG instances—especially from the Station class—consistently improves performance over no enhancement and over other KGs. For LLM-based ridership prediction, RTSKG also outperforms other KGs and the vanilla LLM approach. The paper concludes that RTSKG is effective for city-level rail transit station analysis and has potential applications beyond rail transit to other urban infrastructure systems.

Improvements to AI Systems:

1. Spatial-Semantic Graph Reasoning for Urban Prediction

• Integrate RTSKG’s unified ontology (e.g., Station, Functional Area, Block, POI) into AI models to jointly encode geographic adjacency, intersection, and non-geographic relations.

• Improved system: Predict ridership or store viability by reasoning over multi-hop paths (e.g., Station → Line → Functional Area → POI) instead of treating entities as isolated features.

2. Hierarchical and Cyclic Structure-Aware Embedding

• Adopt the GIE (Graph Isomorphism Embedding) approach, which outperformed 10 other KGE models on RTSKG, to handle both hierarchical (e.g., Borough → Functional Area → Station) and cyclic (e.g., Station ↔ Road ↔ Block) relations.

• Improved system: Learn embeddings that preserve transitivity and symmetry in urban networks, enabling better link prediction for missing station-area connections or new POI placements.

3. Isochrone-Based Contextualization for Station-Area Tasks

• Use RTSKG’s 5/10-minute walking isochrone station areas as a spatial boundary for feature extraction and model input.

• Improved system: Recommend stores or estimate foot traffic by aggregating POI, road, and block data within walkable reach, rather than using arbitrary administrative boundaries—yielding more realistic local demand modeling.

4. Knowledge-Enhanced Ridership Prediction (Traditional Models)

• Inject RTSKG triples (especially from the Station class) as auxiliary features into gradient boosting or neural regressors, replacing flat tabular inputs with graph-structured context.

• Improved system: Achieve consistent ridership forecast gains by learning from station-line-zone interactions (e.g., Station → Line → Functional Area → POI Category) that capture temporal and spatial spillover effects.

5. LLM-Based Reasoning with Structured Knowledge

• Augment large language models (e.g., GPT-4, Llama) with RTSKG subgraphs as retrieval-augmented generation (RAG) context, providing explicit spatial-semantic facts during inference.

• Improved system: Answer complex urban queries (e.g., Which station areas have high retail potential but low transit frequency?) by grounding LLM outputs in graph triples, reducing hallucination and improving interpretability.

6. Transferable Urban Infrastructure Modeling

• Use RTSKG’s ontology and relation categories as a template for constructing similar KGs for other cities or infrastructure (e.g., airports, hospitals, power grids).

• Improved system: A pre-trained graph encoder on RTSKG can be fine-tuned for new urban systems, enabling zero-shot or few-shot prediction of service demand, congestion, or accessibility without retraining from scratch.

7. Multi-Scale Spatial Interaction Simulation

• Leverage the 19 relation types (geographic location, adjacency, intersection, non-geographic) to simulate cascading effects (e.g., a road closure affecting station ridership and nearby POI footfall).

• Improved system: A graph neural network with RTSKG can model dynamic perturbations, supporting real-time urban planning and emergency response by predicting ripple effects across boroughs, blocks, and functional areas.

8. Benchmarking and Evaluation Protocol

• Adopt RTSKG’s two-task evaluation (store recommendation + ridership prediction) as a standard benchmark for urban KG models.

• Improved system: Researchers can compare new AI models against GIE and LLM baselines on a fixed, reproducible dataset, accelerating progress in urban spatial intelligence.

Rail transit systems play a vital role in urban mobility and economic development. As key components of such systems, rail transit stations function as critical transport hubs that enhance urban accessibility and stimulate development in surrounding areas. City-level rail transit station related tasks (e.g., ridership prediction) require large-scale urban data, but current studies often neglect complex interactions among various urban entities in terms of data organization. In this paper, to address the above issue, we build a Rail Transit Station Knowledge Graph (RTSKG) dataset which explicitly models the spatial and semantic interactions among different kinds of urban entities, to benefit city-level rail transit station related tasks. RTSKG integrates heterogeneous urban entities, such as rail transit stations, road segments, and points of interest, with a specially designed unified schema, and is accessible as Linked Data at https://w3id.org/rtskg/. Evaluations on station-area store recommendation and knowledge-enhanced ridership prediction demonstrate the effectiveness of RTSKG, highlighting its potential to support city-level rail transit station analysis.