RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset Shutong Zhu1 , Tianxing Wu1,2(B) , Runfeng Liu1 , Yuang Gu1 , Xuan He1 , and Yuan Zhu3 School of Computer Science and Engineering, Southeast University, China {shutong_zhu,tianxingwu}@seu.edu.cn 2 Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China 3 School of Architecture, Southeast University, China

arXiv:2608.11080v1 [cs.AI] 11 Aug 2026

1

Abstract. 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 citylevel rail transit station analysis. Keywords: Rail Transit Station · Knowledge Graph · Knowledge Graph Embedding.

1

Introduction

Rail transit system is a local rail system providing passenger service within and around urban or suburban areas. As a main form of public transport, rail transit system is characterized by high capacity, high speed, and a high level of safety [16], while also offering advantages in efficiency and punctuality over other modes of transportation like conventional bus systems [47,7,20]. It plays a vital role in urban mobility and economic development since it can decrease congestion-related delay and help the development along transit corridors [39,31]. As key components of rail transit systems, rail transit stations are critical transport hubs which have been widely studied for their impact on urban development [12]. Traditional research on rail transit stations often relies on manually

2

S. Zhu et al.

Station B

POI

Station A Passenger

Passenger

Fig. 1. An example of the interactions among two stations and a POI, implying the ridership of both stations mutually influences each other.

collected data [24] and focuses on a limited number of stations [50], thus fails to capture common features of the stations across the entire rail transit system, and cannot support city-level rail transit station related tasks (e.g., ridership prediction). Consequently, recent studies have shifted toward using large-scale data sources [15,37] to conduct rail transit station analysis. However, they often treat rail transit stations, points of interest (POIs), and etc. as independent entities, neglecting complex interactions among various urban entities, which limits their application in real-world scenarios. For example, as shown in Figure 1, station A and station B are adjacent on the same rail transit line, and the walking times from a certain POI to both stations are similar. Passengers departing from the POI may choose either station for boarding. Therefore, the ridership of both stations mutually influences each other, i.e., an increase in the ridership of station A may probably cause a corresponding decrease in the ridership of station B. Explicitly modeling such interactions among the POI and both stations could benefit the task of ridership prediction. Actually, rail transit station related urban entities are distributed across heterogeneous data sources, and it is non-trivial to integrate different kinds of entities [9] and model their interactions. Recently, knowledge graphs (KGs) have been adopted for organizing urban data due to their ability to model complex relations among urban entities [22,36], but current studies do not involve rail transit stations, which prevents their application in city-level rail transit station related tasks. Thus, in this paper, we propose RTSKG, the Rail Transit Station Knowledge Graph dataset, which explicitly models the spatial and semantic interactions among different kinds of urban entities. Specifically, RTSKG consists of two sub-KGs for New York City and Chicago, respectively. For each city, as shown in the overview of our work (Figure 2), we first collect different kinds of urban data, including rail transit data, administrative data, road data, and POI data. We then build the KG by extracting entities and relations from such collected

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset Urban Data Collection

KG Construction

KG Evaluation RTSKG

Administrative Division Data

Rail Transit Data

3

Ontology Building Station

• •

Rail Transit Stations Rail Transit Lines …

• •

Administrative Boroughs Functional Areas

POI Data

Road Data

•

Road Segments and their Categories …

Station Area

Block

•

Knowledge Graph Embedding

Large Language Model

Station-Area Store Recommendation

Knowledge-Enhanced Ridership Prediction

POI

Fact Extraction

Points of Interest and their Categories

Fig. 2. The overview of our work. Different kinds of urban data are collected to construct RTSKG with a unified schema, and the built KG is evaluated by station-area store recommendation and knowledge-enhanced ridership prediction.

data with a unified schema which models important classes and relations relevant to rail transit systems. We finally conduct experiments on the designed city-level rail transit station related tasks to show the effectiveness of RTSKG. Contributions. The main contributions of this paper are summarized as follows: – We propose RTSKG, the first rail transit station knowledge graph dataset, which explicitly models the spatial and semantic interactions among different urban entities with a specially designed ontology as its unified schema. – We design two representative city-level rail transit station related tasks: station-area store recommendation and knowledge-enhanced ridership prediction, which are used to evaluate the practical utility of RTSKG. – We conduct comprehensive experiments on the designed city-level rail transit station related tasks. The results show that leveraging RTSKG for the above tasks has the best performance compared with using existing urban KGs, reflecting the value of RTSKG in city-level rail transit station analysis.

2

Related Work

2.1

Rail Transit Station Analysis

Rail transit station analysis has been a long-standing research topic. Traditional research often relies on manually collected data and covers a limited number of stations. For example, Wu et al. [41] utilizes manually measured meteorological data for anthropogenic heat estimating, Nafi et al. [21] focuses on Hamad Hospital station in Doha and investigates its influence on the surrounding areas. Such small-scale data cannot support city-level rail transit station related tasks. Recent studies try to use large-scale data sources for rail transit station analysis. For example, Yu et al. [45] collects large-scale POI data to characterize rail transit stations at the city scale. Lee et al. [15] extracts a vast amount of smart

4

S. Zhu et al.

card data from passengers who use rail transit for their trips, and predicts the ridership of rail transit stations based on such data. Zhu et al. [51] constructs the environmental factors of stations based on POI for ridership prediction. However, these data often simply treat urban entities like POIs as independent entities, neglecting complex interactions among heterogeneous urban entities, and we aim to solve this issue in this paper by building a large-scale knowledge graph dataset RTSKG. It explicitly models the spatial and semantic interactions among different kinds of urban entities to facilitate city-level rail transit analysis.

2.2

Resources for Urban Modeling

Urban KGs model relations among urban entities, providing a structured way to integrate and organize such heterogeneous urban data, and have been applied to different urban tasks. For example, among non-public urban KGs, STKG [34] jointly models category information of venues and temporal information, and is used for mobility prediction; Knowsite [18] constructs a KG which captures cities’ key elements and complex relationships for site selection; KG-MUP [19] builds a KG with domain entities and applies it for user profile inference. Recently, several open urban KGs have been built. For example, UUKG [22] is the first open urban KG dataset for knowledge-enhanced urban spatiotemporal predictions, and HUSK [36] further designs a function zone construction method to capture fine-grained semantic information between POIs and areas, enriching the semantic and spatial relationships in urban KGs. However, these urban KGs do not involve rail transit station related information, so they cannot effectively deal with city-level rail transit station related tasks, which can be addressed by our built RTSKG. Alongside these urban KGs developed for urban tasks, several resources provide representations of knowledge about public transport, geography, and urban infrastructure, thereby guiding the modeling of rail transit systems. For example, the NEPTUNE and Passim datasets [25] are represented in RDF format based on specially developed ontologies and published as Linked Data, describing public transport routes and relevant information on transport services, respectively. LinkedGeoData [28] transforms and represents OpenStreetMap data adhering to the RDF data model. GeoKG [11] uses schema and data layers to represent and integrate geographical knowledge from multi-source data. City Infrastructure Ontologies [8] represent city infrastructure assets and their interdependencies in OWL 2. LibCity-Dataset [35] standardizes urban spatial-temporal data storage using atomic files that represent geographic entities and relations. However, these resources either focus on public transport knowledge with limited coverage of other urban entities relevant to rail transit stations, or represent urban knowledge in a generic manner without specifically organizing entities and relations around rail transit systems. RTSKG addresses this gap by integrating different kinds of urban entities under a specially designed schema.

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

3

5

Knowledge Graph Construction

In this section, we introduce different kinds of urban data we collect. We also present our designed ontology and explain how we extract facts from the collected data to build the KG. 3.1

Urban Data Collection

We collect the urban data grouped into four categories for two metropolises: New York City and Chicago. Rail Transit Data. The rail transit data involve the information of key components of rail transit systems and it contains the data of rail transit stations, rail transit lines, and rail transit station entrances & exits. Rail transit stations and rail transit lines are obtained from the Open NY1 and CHI Data Portal2 . We collect station entrances & exits from Open NY and OpenStreetMap3 (OSM) for New York City and Chicago, respectively. Administrative Division Data. Administrative divisions represent hierarchical geographic regions created by governments. We collect two types of administrative divisions, i.e., administrative boroughs and functional areas. Administrative boroughs represent large-scale administrative divisions that support city management and public service delivery. Functional areas refer to the subdivisions of a city according to their functions, such as residential areas. The administrative division data are obtained from NYC Opendata4 and CHI Data Portal. Road Data. Road data involves information about the urban road network including road segments, road categories and roadbeds. We collect road segments and road categories for New York City and Chicago from NYC OpenData and the Chicago Data Portal, respectively. Roadbeds are the regions representing the width of road segments, which can be applied to generate urban entities like urban blocks [27] for analysis. We collect roadbeds directly from NYC OpenData for New York City and generate roadbeds based on self-defined widths of road segments for Chicago. POI Data. POIs are specific places where people gather and conduct daily activities [17] and they are widely used to support city-level urban tasks. The POI data are obtained from the famous urban KG dataset UUKG [22]. 3.2

Ontology Building & Fact Extraction

This subsection introduces the construction process of KG. 1

https://data.ny.gov/ https://data.cityofchicago.org/ 3 https://www.openstreetmap.org/ 4 https://opendata.cityofnewyork.us/ 2

6

S. Zhu et al.

Ontology Building We build the RTSKG ontology as the schema [26] to guide extracting and integrating factual knowledge from different kinds of urban data. Following the broadly accepted procedure [23,40], we first determine the domain and scope of the ontology. RTSKG is designed to model the interactions among different kinds of urban entities and support city-level rail transit station related tasks. Based on concrete needs in urban analysis, we formulate a set of competency questions (listed in our GitHub repository) to specify the important concepts to be covered by the ontology. We then consider reusing existing vocabularies. For example, we reuse the standard RDF, RDFS, OWL and Dublin Core vocabularies, including rdf:type linking from instances to classes, rdfs:label recording the labels of classes, relations (i.e., object properties) and attributes (i.e., data properties), owl:versionIRI identifying the IRI used to identify the version of the ontology, dcterms:description providing a textual description of the ontology, rdfs:domain specifying that any resource that has a given relation or attribute is an instance of one or more classes, and rdfs:range specifying that the values of a relation or attribute are instances of one or more classes. We then define the classes, relations, and attributes in the RTSKG ontology, where attributes are used to describe basic spatiotemporal information of urban entities (e.g., coordinates of POIs). Due to space limitations, this subsection focuses on the classes and relations of RTSKG, while the full ontology is available online: https://w3id.org/rtskg/ontology. Figure 3 shows a part of our specially designed ontology. We define ten classes as follows. (1) Borough refers to administrative borough describing high-level administrative boundaries of a city; (2) Functional Area represents the subdivision of a city. Along with Borough, it enables urban entities from different data sources to be aligned according to administrative divisions; (3) Station refers to the rail transit station, which serves as a transport hub for passengers to board or alight from the rail transit vehicles like subway train; (4) Station Area refers to the region accessible from a rail transit station within a specified walking time, providing a local area around the station for analysis; (5) Line refers to the rail transit line, which is represented as an abstract sequence of connected stations along the route; (6) Road denotes the road segment in a city, which forms the traffic network and provides a basis for human mobility, potentially supporting city-level rail transit station related tasks; (7) Block refers to the urban block, which is a finer-grained spatial unit that constitutes the fundamental element of the physical structure of urban areas and supports urban analysis; (8) POI represents the basic functional units in a city, which are widely used to support city-level urban tasks; (9) POI Category refers to category of POI, which characterizes the function of POI; (10) Road Category is the category determined by the hierarchy or function of the road. We also define nineteen relations, which are further categorized into four types as follows. Geographic Location. This relation type models that one instance locates at another instance. We define seven geographic location relations in RTSKG. Specifically, poi_locates_at_station_area and poi_locates_at_block rep-

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

7

resent location relations between point and polygonal instances. station_ar ea_locates_at_borough, station_area_locates_at_functional_area, block_locates_at_borough, block_locates_at_functional_area and functional_area_locates_at_borough represent location relations between two polygonal instances. Geographic Adjacency. This relation type models that one instance is adjacent to another instance, and we define two geographic adjacency relations in RTSKG: borough_is_adjacent_to_borough and functional_area_is_adjacent_to _functional_area. Both relations enable the ontology to model the adjacency between two polygonal instances. Geographic Intersection. This relation type models that one instance intersects another instance, and we define six geographic intersection relations in RTSKG. Specifically, road_intersects_borough, road_intersects_functional_area, and road_intersects_station_area model the intersections between linear and polygonal instances, characterizing the spatial distribution of the road network across geographic regions. station_area_intersects_block, station_ area_intersects_station_area, and block_intersects_station_area represent intersections between two polygonal instances, capturing potential interactions through their shared spatial regions (e.g., the region of overlap between two station areas). Non-Geographic Relation. This relation type covers the relations other than the above three relation types. We define four non-geographic relations as follows: station_area_belongs_to_station, station_belongs_to_line, poi_has_ poi_category, and road_has_road_category. These relations do not capture spatial interactions between instances, but are still important in rail transit station analysis. Fact Extraction We extract facts from the data collected in Section 3.1. To handle heterogeneous urban data, we employ the GeoPandas [13], a Python library for dealing with geospatial data, to extract urban entities and facts. Specifically, we use GeoPandas to extract the instances belonging to the classes Borough, Functional Area, Station, Road, and POI, together with their spatial locations, from the geospatial data in the shapefile format. We also obtain the instances belonging to the classes Line, POI Category, and Road Category directly from the tabular data we collect with the comma-separated values format. For the instances in the class Station Area, following the definition in [33], we construct a station area for each station based on the isochrone regions (i.e., the region can be reached within a certain walking time) of its entrances or exits. For each entrance or exit, we generate an isochrone region using the Mapbox Isochrone API5 based on its geographic coordinates. The station area of a station is then obtained by taking the union of the isochrone regions of all the entrances 5

https://www.mapbox.com/

8

S. Zhu et al.

Fig. 3. A part of the RTSKG ontology.

and exits. In this paper, we construct station areas with two fixed walking times, i.e., 5 minutes and 10 minutes, which are common thresholds in transit-oriented development (TOD) and station-area studies [29,6], for each station. Figure 4 (a) shows the station areas of the 18th Street station with 5-minute and 10minute walking times. For the instances in the class Block, we define them as the regions in the city excluding roadbeds. Figure 4 (b) visualizes a part of blocks in New York City. After obtaining the instances of all classes, we use GeoPandas to extract all relations of geographic location, geographic adjacency, and geographic intersection. For non-geographic relations, we directly obtain them from the collected data with the comma-separated values format. By integrating the extracted facts, we build KGs for New York City and Chicago, respectively. Table 1 shows the statistics of the KGs for the cities over the classes, relations, and attributes defined in the RTSKG ontology, while the dumps released on Zenodo additionally provide triples with predicates from standard vocabularies. The KG construction process can be extended to other cities by collecting the required urban data and adapting the Python scripts to handle differences in data formats. All class URIs in the namespace https: //w3id.org/rtskg/ontology/class/ and all instance URIs in the namespaces https://w3id.org/rtskg/new_york_city/instance/ and https://w3id.o rg/rtskg/chicago/instance/ are dereferenceable. Table 1. The statistics of RTSKG. City New York City Chicago

#Instances #Relations #Relation Triples #Attribute Triples 238,839 151,845

19 19

1,113,638 448,716

239,090 151,943

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset station area with 5-minute walking time

entrance/exit of 18th Street Station

station area with 10-minute walking time

block

(a)

9

(b)

Fig. 4. An example of station areas and blocks in New York City. (a) Station areas of 18th Street station, where the inner polygon represents the station area with 5-minute walking time, the outer polygon represents the station area with 10-minute walking time, and the nodes represent the entrances or exits of 18th Street Station. (b) A part of Blocks.

4

Knowledge Graph Embedding

In this section, we apply knowledge graph embedding (KGE) models to RTSKG and select the best KGE model on link prediction for city-level rail transit station related tasks (details will be given in Section 5 and Section 6). Problem Formulation. Given a KG G, KGE is to learn vector representations (i.e., embeddings) for instances and relations in a dense and low-dimensional space. The embedding of the instance i ∈ I is denoted as i and the embedding of the relation r ∈ R is denoted as r, where I and R are respectively the sets of instances and relations.

4.1

Experimental Setup

We compared the performance of different KGE models on link prediction which aims to predict the missing tail or head instance for a query (h, r, ?) or (?, r, t), where h, t ∈ I denote the head and tail instances, respectively. Evaluation Protocol. We reported the following evaluation metrics for link prediction: (1) Hits@k (k=1,3), the percentage of correct instances in the top-k ranked instances. (2) Mean reciprocal rank (MRR), the average multiplicative inverse of the ranks for all instances.

10

S. Zhu et al.

Models. We employed various KGE models to represent KGs, and chose the best model as the base model for city-level rail transit station related tasks. The models are listed as follows: (1) TransE [3] is the classic translation-based KGE model. (2) DistMult [43] is a bilinear KGE model with tensor decomposition. (3) ComplEx [32] is an extension of DistMult which learns KG embeddings in a complex space. (4) RotatE [30] is a KGE model which defines each relation as a rotation in the complex vector space. (5) MuRE [1] is a translational distance KGE model with a diagonal relational matrix. (6) TuckER [2] is a KGE model based on tucker decomposition of the binary tensor representation of triples. (7) QuatE [46] is a KGE model which introduces more expressive hypercomplex representations to model entities and relations. (8) AttH [5] is a KGE model which combines hyperbolic reflections and rotations with attention to learn embeddings in KG. (9) RefH/RotH [5] are variants of AttH, and they are hyperbolic KGE models with reflection and rotation, respectively. (10) GIE [4] is a product space KGE model, which aims to learn spatial structures interactively between the Euclidean, hyperbolic and hyperspherical spaces. Implementation Details. All models were implemented by PyTorch, and all experiments were conducted on an RTX 3090 GPU. Following the previous work [22], we fixed the negative sampling size as 50 and the embedding dimension as 32 for all models. 4.2

Results

Table 2 reports the link prediction results in the sub-KGs for New York City and Chicago. Experimental results indicate that GIE outperforms other KGE models across all link prediction metrics on both sub-KGs. This is mainly because our ontology contains both hierarchical structures and cyclic structures. As shown in Figure 3, the relations among Station Area, Functional Area, and Borough form a hierarchical structure, and the relations among Station Area, POI, and Block form a cyclic structure. GIE is designed to capture such structures simultaneously [22]. Motivated by GIE’s superior KGE ability, we chose GIE as the base model to generate instance and relation embeddings for city-level rail transit station related tasks.

5

Station-Area Store Recommendation

In this section, we introduce the task of station-area store recommendation, which aims to evaluate the utility of our RTSKG for the analysis of economic development in the surrounding areas of stations. Problem Formulation. Station-area store recommendation can be formally modeled as a special link prediction task. Given a POI instance hp located in a station area, station-area store recommendation aims to predict the category through the query (hp , rp , ?), where rp denotes the relation poi_has_poi_category.

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

11

Table 2. Link prediction results of eleven KGE models on the sub-KGs for New York City and Chicago. Model

5.1

New York City Chicago Hits@1 Hits@3 MRR Hits@1 Hits@3 MRR

TransE DistMult MuRE TuckER RotatE ComplEx QuatE RotH RefH ATTH

0.291 0.302 0.325 0.200 0.140 0.208 0.291 0.304 0.304 0.319

0.417 0.450 0.460 0.351 0.304 0.384 0.443 0.439 0.435 0.463

0.378 0.389 0.413 0.297 0.250 0.315 0.381 0.395 0.394 0.412

0.272 0.196 0.308 0.185 0.079 0.154 0.237 0.277 0.277 0.310

0.394 0.349 0.430 0.316 0.190 0.305 0.377 0.404 0.401 0.432

0.355 0.289 0.388 0.272 0.169 0.250 0.322 0.365 0.363 0.392

GIE

0.326

0.467

0.417

0.317

0.438

0.397

Experimental Setup

We conducted station-area store recommendation in New York City and Chicago. We still selected MRR and Hits@k (k=1,3) as the evaluation metrics. We compared RTSKG with two existing open urban KG datasets, i.e., UUKG [22] and HUSK [36]. Data Description. For each city, we sampled POIs located at station areas with 5minute walking time, and ensured that each sampled POI appears as an instance in each of the three KG datasets (i.e., RTSKG, UUKG, and HUSK). We then constructed the validation and test sets by extracting relation triples with the relation poi_has_poi_category whose head instances are these sampled POIs. All remaining relation triples were used as the training set for each KG dataset. Implementation Details. We randomly sampled 2,500 relation triples and split them into 500 relation triples for validation and 2,000 relation triples for test. We employed GIE as the base model, and we adopted the same hyper-parameter settings as in Section 4.1.

5.2

Results

As shown in Table 3, GIE achieves the best Hits@k and MRR when leveraging RTSKG, compared with leveraging other KGs. By performing the task of stationarea store recommendation with RTSKG, we can more accurately identify the most suitable development directions for stores in station areas, which enables more efficient resource allocation and contributes to the economic development of the surrounding areas of the stations.

12

S. Zhu et al. Table 3. The results on station-area store recommendation. KG

6

New York City Chicago Hits@1 Hits@3 MRR Hits@1 Hits@3 MRR

UUKG HUSK

0.698 0.705

0.775 0.785

0.757 0.770

0.264 0.356

0.533 0.615

0.455 0.524

RTSKG

0.720

0.794

0.782

0.380

0.694

0.570

Knowledge-Enhanced Ridership Prediction

In this section, we introduce the task of knowledge-enhanced ridership prediction, which aims to evaluate the practical utility of our RTSKG for the analysis of rail transit station related urban mobility. Problem Formulation. Ridership prediction is a classic urban task, and it is typically formulated as time series prediction, which aims to predict the station ridership on the next day based on the past station ridership sequence. Knowledge-enhanced ridership prediction aims to perform ridership prediction by incorporating the KG for knowledge enhancement. Let X (t) ∈ Ra denote the ridership observations at the time step t, where a denotes the number of rail transit stations. Given the past observations (X (t−T ) , X (t−T +1) , . . . , X (t−1) ), and an external KG G, where T is the input length of the past observations, knowledge-enhanced ridership prediction aims
to employ a model m to predict X (t) = m (X (t−T ) , X (t−T +1) , . . . , X (t−1) ), G , where m(·) is the function that model m transforms a sequence of ridership to a single ridership value. Depending on the type of model adopted, knowledge-enhanced ridership prediction can be further categorized into traditional knowledge-enhanced ridership prediction and LLM-based knowledge-enhanced ridership prediction. Traditional Knowledge-Enhanced Ridership Prediction. Traditional knowledgeenhanced ridership prediction [22,36] typically performs knowledge enhancement by using KG embeddings as external knowledge features and directly concatenating them with past ridership observations as the input of the model m. LLM-based Knowledge-Enhanced Ridership Prediction. Recently, large language models (LLMs) have shown impressive performance on various tasks, and applications combining LLMs with KGs have become a significant research direction as well. Motivated by this, we followed the previous work [42] and proposed the task of LLM-based knowledge-enhanced ridership prediction. Specifically, we verbalized each time series instance (i.e., the ridership observations over the past days of a station) into a natural language query, e.g., “From September 05 to September 16, 2025, the rail transit station ‘Myrtle-Wyckoff Avs’ located at the ‘Bushwick North’ functional area recorded the following ridership numbers: 18030 12496 ... 18461 on each day. What is the ridership going to be on September 17, 2025?”. Meanwhile, we verbalized each KG relation triple into a

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

13

natural language statement, e.g., “The rail transit station ‘Myrtle-Wyckoff Avs’ belongs to the rail transit line ‘Canarsie” ’. We then encoded all the queries and verbalized relation triples as embeddings with a pretrained language model, and computed similarity scores between the queries and relation triples by cosine similarities of their embeddings. For each query, we retrieved the top-n most relevant verbalized relation triples from the KG based on these similarity scores. Next, we used an LLM to summarize the retrieved relation triples into an abstract, which serves as external knowledge for the query. Finally, we concatenated the query with the abstract to form a prompt and fed it into another LLM to perform knowledge-enhanced ridership prediction. The prompts are provided in our GitHub repository.

6.1

Experimental Setup

We conducted knowledge-enhanced ridership prediction in New York City and Chicago. Evaluation Protocol. For knowledge-enhanced ridership prediction, we chose evaluation metrics as follows: Mean Absolute Error (MAE): the mean absolute error of predicted ridership and ground ridership, and Root Mean Squared Error (RMSE): the root mean squared error of predicted ridership and ground ridership. Data Description. We collected daily ridership records of rail transit stations from U.S. Government’s Open Data6 , and we manually filtered out the records that may be ambiguous or erroneous. The statistics of filtered ridership data are summarized in Table 4. Models. For traditional knowledge-enhanced ridership prediction, we employed five representative spatiotemporal prediction models: (1) STGCN [44] is spatiotemporal graph convolutional network, which has complete convolutional structures. (2) ASTGCN [10] is attention based spatial-temporal graph convolutional network, which combines the spatial-temporal attention mechanism and the spatial-temporal convolution. (3) TGCN [48] is a graph convolutional network, which is combined with the gated recurrent unit. (4) MSTGCN [10] is multi-component spatial-temporal graph convolution network, which is a variant of ASTGCN. (5) STPGCN [49] is spatiotemporal position-aware graph convolutional network, which constructs a trainable embedding module and represents the spatial and temporal positions of the nodes in graph. For LLMbased knowledge-enhanced ridership prediction, we adopted GPT-4o7 , a widely adopted LLM provided by OpenAI, as the base LLM. 6 7

https://data.gov/ https://openai.com/index/hello-gpt-4o/

14

S. Zhu et al. Table 4. The statistics of ridership data City

New York City

Chicago

# of records

343368

114456

Time span

01/01/2023-09/30/2025

Baselines. We compared the models enhanced by RTSKG with the variants without knowledge enhancement (denoted as Vanilla). We also evaluated two existing urban KGs, UUKG and HUSK, by leveraging them for knowledge enhancement. For traditional knowledge-enhanced ridership prediction, since neither existing KG contains rail transit stations, we applied their instances from commonly adopted class (i.e., Functional Area) for knowledge enhancement, denoted as w/UUKG and w/HUSK. For fair comparison, we additionally enhanced the model with instances in the class Station and the class Functional Area from RTSKG, denoted as w/RTSKG-S and w/RTSKG-F, respectively. For LLM-based knowledge-enhanced ridership prediction, we directly used all the KGs as external knowledge bases for knowledge enhancement. The variant without knowledge enhancement is denoted as Vanilla-LLM, and the variants enhanced with RTSKG, HUSK, and UUKG are denoted as w/RTSKG-LLM, w/HUSK-LLM, and w/UUKG-LLM, respectively. Implementation Details. In this task, we used the observations from the past 12 days to predict ridership for the next day. For traditional knowledge-enhanced ridership prediction, we split the dataset into training, validation, and test sets with a ratio of 7:1:2. All models were trained using Adam [14] with a batch size of 64. For LLM-based knowledge-enhanced ridership prediction, we randomly sampled 500 records from the test set for evaluation. We set the number of retrieved relation triples n as 50 and we adopted MiniLM [38] to encode queries and verbalized relation triples as embeddings. 6.2

Results

Traditional Knowledge-Enhanced Ridership Prediction. As shown in Table 5, the experimental results indicate that knowledge enhancement with instances in either the Station class or the Functional Area class from RTSKG consistently improves the performance of models on ridership prediction compared to the models without enhancement. Furthermore, the models enhanced with the instances in the class Station outperform those enhanced with the instances in the class Functional Area, suggesting that explicitly modeling stations in the KG can capture the mutual influence between the ridership of different stations, which is more beneficial for ridership prediction. Additionally, the model enhanced with the instances in the class Functional Area from RTSKG achieves better performance compared with those enhanced by the instances in the class Functional Area from other KGs, as RTSKG models the interactions between different kinds of urban entities, allowing the embeddings of instances in the class Functional

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

15

Table 5. The comparison results of traditional knowledge-enhanced ridership prediction. The best results are indicated by bold numbers. Model

Metric

STGCN

Vanilla w/UUKG w/HUSK w/RTSKG-F w/RTSKG-S

Vanilla w/UUKG w/HUSK ASTGCN w/RTSKG-F w/RTSKG-S

New York City MAE

RMSE

Chicago MAE

RMSE

1490.229 2555.250 272.275 589.112 1196.467 2301.635 259.893 546.069 1193.374 2322.347 250.546 544.304 936.359 1819.504 247.839 543.506 735.471 1563.717 241.169 538.621 101.066 94.583 100.330 91.003 70.766

218.336 209.423 213.399 181.706 173.516

61.038 160.673 47.208 136.242 48.003 137.948 44.958 131.662 42.143 126.379

Vanilla 208.735 w/UUKG 217.124 206.557 MSTGCN w/HUSK w/RTSKG-F 193.388 w/RTSKG-S 176.531

404.118 389.656 374.751 366.905 353.559

87.532 185.724 85.472 179.737 83.794 182.502 81.437 179.197 81.330 178.617

TGCN

Vanilla 3724.667 6686.524 404.930 719.092 w/UUKG 3680.524 6069.055 399.700 717.116 w/HUSK 3619.817 6242.119 393.618 698.174 w/RTSKG-F 3383.251 6004.507 381.304 693.866 w/RTSKG-S 3222.358 5580.707 378.903 687.910

Vanilla 1100.163 2304.365 319.826 695.963 w/UUKG 1081.648 2170.809 319.722 681.038 w/HUSK 1088.856 2238.719 315.819 680.975 STPGCN w/RTSKG-F 1050.264 2120.993 312.528 678.507 w/RTSKG-S 1007.729 2075.475 307.396 673.780

Area to implicitly include features from stations. The experimental results fully demonstrate the effectiveness of RTSKG, highlighting that explicitly modeling rail transit stations in the KG helps capture the implicit interactions among urban entities, and helps government departments adjust operations in advance of ridership changes, thus can benefit the analysis of urban mobility. LLM-Based Knowledge-Enhanced Ridership Prediction. Figure 5 illustrates the comparison results of LLM-based knowledge-enhanced ridership prediction. Experimental results show that our RTSKG, as an external knowledge base for knowledge enhancement, achieves better performance than directly applying LLMs for ridership prediction and consistently outperforms leveraging other KGs for knowledge enhancement. This demonstrates that RTSKG is a high-quality dataset for city-level rail transit station analysis, and it maintains effectiveness even in the rapidly evolving field of LLM applications. RTSKG holds significant potential for supporting urban developments. In the future, RTSKG can be used

16

S. Zhu et al.

Fig. 5. The comparison results of LLM-based knowledge-enhanced ridership prediction.

to support more LLM-based applications, such as city-level rail transit station related Q&A.

7

Conclusion

In this paper, we propose RTSKG, a new rail transit station knowledge graph dataset, which is designed to explicitly model rail transit stations and the complex spatial and semantic interactions among different kinds of urban entities for city-level rail transit station related tasks. Experimental results demonstrate that leveraging RTSKG for station-area store recommendation and knowledgeenhanced ridership prediction has the best performance compared with leveraging other urban KGs. By providing a reusable dataset for city-level rail transit station related tasks, our study has potential applications not only in rail transit systems but also in other urban infrastructure systems, which contributes to enhancing urban development. Acknowledgments. This work is supported by the NSFC (Grant No. 52378009, 62376058), ZhiShan Young Scholar Program of Southeast University (Grant No. 2242026RCB0011), the Fundamental Research Funds for the Central Universities, the Southeast University Interdisciplinary Research Program for Young Scholars, and the Big Data Computing Center of Southeast University. Resource Availability Statement: All classes, relations and instances in RTSKG are identified by permanent dereferenceable URIs in w3id8 . All data are available as RDF dump files on Zenodo9 , and the basic information and source code of the RTSKG project can be accessed at GitHub10 . The data resources are released under the CC BY-SA 4.0 License, and the source code is released under the Apache License 2.0. 8

https://w3id.org/rtskg https://zenodo.org/records/21778005 10 https://github.com/seucoin/RTSKG 9

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

17

Declaration of Use of Generative AI During the preparation of this work, generative AI tools were used exclusively for providing ideas of creating illustrative figures. The authors reviewed and approved all figures and take full responsibility for the content of the publication.

References 1. Balazevic, I., Allen, C., Hospedales, T.M.: Multi-relational poincaré graph embeddings. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d’Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 4465–4475 (2019), https://proceedings.neurips.cc/paper/2019/hash/f8b932c70d0 b2e6bf071729a4fa68dfc-Abstract.html 2. Balazevic, I., Allen, C., Hospedales, T.M.: Tucker: Tensor factorization for knowledge graph completion. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLPIJCNLP 2019, Hong Kong, China, November 3-7, 2019. pp. 5184–5193. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/D19-1522 3. Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O.: Translating embeddings for modeling multi-relational data. In: Burges, C., Bottou, L., Welling, M., Ghahramani, Z., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 26. Curran Associates, Inc. (2013), https: //proceedings.neurips.cc/paper_files/paper/2013/file/1cecc7a77928ca8 133fa24680a88d2f9-Paper.pdf 4. Cao, Z., Xu, Q., Yang, Z., Cao, X., Huang, Q.: Geometry interaction knowledge graph embeddings. In: Thirty-Sixth AAAI Conference on Artificial Intelligence, February 22 - March 1, 2022. pp. 5521–5529. AAAI Press (2022). https://doi.or g/10.1609/AAAI.V36I5.20491 5. Chami, I., Wolf, A., Juan, D., Sala, F., Ravi, S., Ré, C.: Low-dimensional hyperbolic knowledge graph embeddings. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J.R. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020. pp. 6901–6914. Association for Computational Linguistics (2020). https://doi.org/10.18653/V 1/2020.ACL-MAIN.617 6. Curtis, C.: Evolution of the transit-oriented development model for low-density cities: A case study of perth’s new railway corridor. Planning Practice & Research 23(3), 285–302 (2008). https://doi.org/10.1080/02697450802423559 7. Ding, H., Di, Y., Zheng, X., Liu, K., Zhang, W., Zheng, L.: Passenger arrival distribution model and riding guidance on an urban rail transit platform. Physica A: Statistical Mechanics and its Applications 571, 125847 (2021). https://doi. org/10.1016/j.physa.2021.125847 8. Du, H., Wei, L., Dimitrova, V., Magee, D.R., Clarke, B., Collins, R., Entwisle, D., Torbaghan, M.E., Curioni, G., Stirling, R., Reeves, H., Cohn, A.G.: City infrastructure ontologies. Comput. Environ. Urban Syst. 104, 101991 (2023). https://doi.org/10.1016/J.COMPENVURBSYS.2023.101991

18

S. Zhu et al.

9. Dussin Bampi, M., Ourique de Morais, W., Olszewska, J.I., Pignaton De Freitas, E.: Ontology-driven monitoring system for ambient assisted living. The Knowledge Engineering Review 40, e2 (2025). https://doi.org/10.1017/S0269888925000 037 10. Guo, S., Lin, Y., Feng, N., Song, C., Wan, H.: Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In: The Thirty-Third AAAI Conference on Artificial Intelligence, Honolulu, Hawaii, USA, January 27 February 1, 2019. pp. 922–929. AAAI Press (2019). https://doi.org/10.1609/ AAAI.V33I01.3301922 11. Guo, X., Qian, H., Wu, F., Liu, J.: A method for constructing geographical knowledge graph from multisource data. Sustainability 13(19) (2021). https: //doi.org/10.3390/su131910602 12. Jiang, H., Liu, R., Luo, S., Yi, D., Zhang, J.: Understanding metro station areas’ functional characteristics via embedding representation: A case study of shanghai. Scientific Reports 15(1), 2725 (Jan 2025). https://doi.org/10.1038/s41598-0 25-87336-6 13. Jordahl, K., den Bossche, J.V., Fleischmann, M., Wasserman, J., McBride, J., Gerard, J., Tratner, J., Perry, M., Badaracco, A.G., Farmer, C., Hjelle, G.A., Snow, A.D., Cochran, M., Gillies, S., Culbertson, L., Bartos, M., Eubank, N., maxalbert, Bilogur, A., Rey, S., Ren, C., Arribas-Bel, D., Wasser, L., Wolf, L.J., Journois, M., Wilson, J., Greenhall, A., Holdgraf, C., Filipe, Leblanc, F.: geopandas/geopandas: v0.8.1 (Jul 2020). https://doi.org/10.5281/zenodo.3946761 14. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: International Conference on Learning Representations (ICLR) (2015), https://arxiv.or g/abs/1412.6980 15. Lee, M., Jeon, I., Jun, C.: A deterministic methodology using smart card data for prediction of ridership on public transport. Applied Sciences 12(8) (2022). https://doi.org/10.3390/app12083867 16. Lin, D., Broere, W., Cui, J.: Metro systems and urban development: Impacts and implications. Tunnelling and Underground Space Technology 125, 104509 (2022). https://doi.org/10.1016/j.tust.2022.104509 17. Liu, K., Yin, L., Lu, F., Mou, N.: Visualizing and exploring poi configurations of urban regions on poi-type semantic space. Cities 99, 102610 (2020). https: //doi.org/10.1016/j.cities.2020.102610 18. Liu, Y., Ding, J., Li, Y.: Knowsite: Leveraging urban knowledge graph for site selection. In: Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems. SIGSPATIAL ’23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3589132.36 25640 19. Liu, Y., Zhou, Z., Li, Y., Jin, D.: Urban knowledge graph aided mobile user profiling. ACM Trans. Knowl. Discov. Data 18(1) (Oct 2023). https://doi.org/10 .1145/3596604 20. Luo, J., Zhou, D., Ma, W., Zhao, G.: Analyzing rail traffic diversion based on machine learning technique considering transportation security. Journal of Advanced Transportation 2022(1), 8349173 (2022). https://doi.org/10.1155/2022/83491 73 21. Nafi, S., Furlan, R., Grosvald, M., Al-Matwi, R., Lifam Marthya, K.: Transitoriented development in doha: The case of the al sadd neighborhood and hamad hospital metro station. Designs 5(4) (2021). https://doi.org/10.3390/designs5 040061

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

19

22. Ning, Y., Liu, H., Wang, H., Zeng, Z., Xiong, H.: Uukg: Unified urban knowledge graph dataset for urban spatiotemporal prediction. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 62442–62456. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/c4a 30a4dd840cfeff30ba4d2661ff097-Paper-Datasets_and_Benchmarks.pdf 23. Noy, N.F., McGuinness, D.L.: Ontology development 101: A guide to creating your first ontology. Tech. rep. (march 2001), https://protege.stanford.edu/publica tions/ontology_development/ontology101-noy-mcguinness.html 24. Pan, C.: Research on land use optimization around rail transit stations under tod model——taking yishan road station of tianjin metro line 8 as an example. Intelligent Building & City Information (02), 33–35 (2023). https://doi.org/10 .13655/j.cnki.ibci.2023.02.010 25. Plu, J., Scharffe, F.: Publishing and linking transport data on the web: extended version. In: Raschia, G., Theobald, M. (eds.) International Workshop on Open Data, WOD ’12, Nantes, France, May 25, 2012. pp. 62–69. ACM (2012). https: //doi.org/10.1145/2422604.2422614 26. Rashid, S.M., McGuinness, D.L.: Designing and evaluating an ensemble reasoningbased clinical decision support system. Data Intelligence 7(1), 1–39 (2025). https: //doi.org/10.3724/2096-7004.di.2025.0001 27. Shakibamanesh, A., Ebrahimi, B.: Toward practical criteria for analyzing and designing urban blocks. In: Almusaed, A., Almssad, A., Hong, L.T. (eds.) Sustainability in Urban Planning and Design, chap. 10. IntechOpen, Rijeka (2020). https://doi.org/10.5772/intechopen.90504 28. Stadler, C., Lehmann, J., Höffner, K., Auer, S.: Linkedgeodata: A core for a web of spatial open data. Semantic Web 3(4), 333–354 (2012). https://doi.org/10.3 233/SW-2011-0052 29. Sun, Y., Han, B., Lu, F.: An overview of tod level assessment around rail transit stations. Urban Rail Transit 10, 1–12 (2024). https://doi.org/10.1007/s40864 -023-00211-3 30. Sun, Z., Deng, Z., Nie, J., Tang, J.: Rotate: Knowledge graph embedding by relational rotation in complex space. In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net (2019), https://openreview.net/forum?id=HkgEQnRqYQ 31. To, W.M.: Centrality of an Urban Rail System. Urban Rail Transit 1(4), 249–256 (Dec 2015). https://doi.org/10.1007/s40864-016-0031-3 32. Trouillon, T., Welbl, J., Riedel, S., Gaussier, E., Bouchard, G.: Complex embeddings for simple link prediction. In: Balcan, M.F., Weinberger, K.Q. (eds.) Proceedings of The 33rd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 48, pp. 2071–2080. PMLR, New York, New York, USA (20–22 Jun 2016), https://proceedings.mlr.press/v48/trouillon16.html 33. Tyndall, J.: Complementarity of dockless mircomobility and rail transit. Journal of Transport Geography 103, 103411 (2022). https://doi.org/10.1016/j.jtra ngeo.2022.103411 34. Wang, H., Yu, Q., Liu, Y., Jin, D., Li, Y.: Spatio-temporal urban knowledge graph enabled mobility prediction. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 5(4) (Dec 2022). https://doi.org/10.1145/3494993 35. Wang, J., Jiang, W., Jiang, J.: Libcity-dataset: a standardized and comprehensive dataset for urban spatial–temporal data mining. Intelligent Transportation Infrastructure 2, liad021 (05 2023). https://doi.org/10.1093/iti/liad021

20

S. Zhu et al.

36. Wang, Q., Wang, G., Pan, Y., Lin, Z., Li, H., Liu, Q., Zhao, K.: Husk: A hierarchically structured urban knowledge graph dataset for multi-level spatial tasks. In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management. p. 6549–6553. CIKM ’25, Association for Computing Machinery, New York, NY, USA (2025). https://doi.org/10.1145/3746252.3761638 37. Wang, Q., Ma, Z., Yang, X., Chien, S.I.J., Zhang, S., Yin, Y.: Exploring spatiotemporal dynamic of metro ridership and the influence of built environment factors at the station level: A case study of nanjing, china. Journal of Transport Geography 129, 104440 (2025). https://doi.org/10.1016/j.jtrangeo.2025.104440 38. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 5776–5788. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file /3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf 39. Weisbrod, G., Reno, A.: Economic impact of public transportation investment. American public transportation association Washington, DC (2009), https://on linepubs.trb.org/onlinepubs/tcrp/docs/TCRPJ-11Task7-FR.pdf 40. Wu, T., Cao, X., Zhu, Y., Wu, F., Gong, T., Wang, Y., Jing, S.: Asdkb: A chinese knowledge base for the early screening and diagnosis of autism spectrum disorder. In: Payne, T.R., Presutti, V., Qi, G., Poveda-Villalón, M., Stoilos, G., Hollink, L., Kaoudi, Z., Cheng, G., Li, J. (eds.) The Semantic Web – ISWC 2023. pp. 59–75. Springer Nature Switzerland, Cham (2023). https://doi.org/10.1007/978-3-0 31-47243-5_4 41. Wu, Y., Shan, Y., Zhou, S., Lai, Y., Xiao, J.: Estimating anthropogenic heat from an urban rail transit station: A case study of qingsheng metro station, guangzhou, china. Sustainable Cities and Society 82, 103895 (2022). https://doi.org/10.1 016/j.scs.2022.103895 42. Xue, H., Salim, F.D.: Promptcast: A new prompt-based learning paradigm for time series forecasting. IEEE Transactions on Knowledge and Data Engineering 36(11), 6851–6864 (2024). https://doi.org/10.1109/TKDE.2023.3342137 43. Yang, B., Yih, S.W.t., He, X., Gao, J., Deng, L.: Embedding entities and relations for learning and inference in knowledge bases. In: Proceedings of the International Conference on Learning Representations (ICLR) 2015 (May 2015), https://www. microsoft.com/en-us/research/publication/embedding-entities-and-relat ions-for-learning-and-inference-in-knowledge-bases/ 44. Yu, B., Yin, H., Zhu, Z.: Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In: Lang, J. (ed.) Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden. pp. 3634–3640. ijcai.org (2018). https: //doi.org/10.24963/IJCAI.2018/505 45. Yu, Z., Zhu, X., Liu, X.: Characterizing metro stations via urban function: Thematic evidence from transit-oriented development (tod) in hong kong. Journal of Transport Geography 99, 103299 (2022). https://doi.org/10.1016/j.jtrangeo .2022.103299 46. Zhang, S., Tay, Y., Yao, L., Liu, Q.: Quaternion knowledge graph embeddings. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d’Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 2731–2741 (2019), https:

RTSKG: Building a Rail Transit Station Knowledge Graph Dataset

21

//proceedings.neurips.cc/paper/2019/hash/d961e9f236177d65d21100592 edb0769-Abstract.html 47. Zhang, Y., Yang, J., Wang, J., Zhao, Y.: Dynamic characteristics of urban rail train in multivehicle marshaling under traction conditions. Applied Sciences 13(5) (2023). https://doi.org/10.3390/app13053022 48. Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., Deng, M., Li, H.: T-GCN: A temporal graph convolutional network for traffic prediction. IEEE Trans. Intell. Transp. Syst. 21(9), 3848–3858 (2020). https://doi.org/10.1109/TITS.2019.29 35152 49. Zhao, Y., Lin, Y., Wen, H., Wei, T., Jin, X., Wan, H.: Spatial-temporal positionaware graph convolution networks for traffic flow forecasting. IEEE Transactions on Intelligent Transportation Systems 24(8), 8650–8666 (2023). https://doi.or g/10.1109/TITS.2022.3220089 50. Zhu, Y., Li, H.: Tod-led urban evolution:an analysis of the renewal strategy of london king’s cross. Urban Environment Design (01), 323–330 (2022). https://do i.org/10.19974/j.cnki.CN21-1508/TU.2022.02.0323 51. Zhu, Z., Zhang, Y., Qiu, S., Zhao, Y., Ma, J., He, Z.: Ridership prediction of urban rail transit stations based on afc and poi data. Journal of Transportation Engineering, Part A: Systems 149(9), 04023077 (2023). https://doi.org/10.106 1/JTEPBS.TEENG-7808

Related documents

• Research on Artificial Intelligence in Rail Pass Management Systems#78695
• Train traffic on the rail network - Türkiye (2020)#816797
• Train traffic on the rail network - Türkiye (2015)#816738
• Transitive reasoning as linear classification#405061
• Buses and coaches traffic performance registered in the reporting country by transport covarage#817425