Multi-Scale Heterogeneous Text-Attributed Graph Datasets From Diverse Domains

• Multi Scales. Our HTAG datasets span multiple scales, ranging from small (24K nodes, 104K edges) to large (5.6M nodes, 29.8M edges). Smaller datasets are suitable for testing computationally intensive algorithms, while larger datasets, such as DBLP and Patent, support the development of scalable models that leverage mini-batching and distributed training.
• Diverse Domains. Our HTAG datasets include heterogeneous graphs that are representative of a wide range of domains: movie collaboration, community question answering, academic, book publication, and patent application. The broad coverage of domains empowers the development and demonstration of graph foundation models and helps differentiate them from domain-specific approaches.
• Realistic and Reproducible Evaluation. We provide an automated evaluation pipeline for HTAGs that streamlines data processing, loading, and model evaluation, ensuring seamless reproducibility. Additionally, we employ time-based data splits for each dataset, which offer a more realistic and meaningful evaluation compared to traditional random splits.
• Open-source Code for Dataset Construction. We have released the complete code for constructing our HTAG datasets, allowing researchers to build larger and more complex heterogeneous text-attribute graph datasets. For example, the CroVal dataset construction code can be used to create web-scale community question-answering networks, such as those derived from StackExchange data dumps. This initiative aims to further advance the field by providing the tools necessary for replicating and extending our datasets for a wide range of applications.

Datasets

The HTAG datasets and dataset construction code are available at HTAG · Datasets on Hugging Face. Each dataset folder contains a README.md file (providing a dataset description and instructions for reproducing dataset construction), a .pkl file (containing the heterogeneous graph information), a .csv or .csv.zip file (with raw text), and a .ipynb file (for constructing the dataset).

Download

from huggingface_hub import snapshot_download

# Download all
snapshot_download(repo_id="Cloudy1225/HTAG", repo_type="dataset", local_dir="./data")

# Or just download heterogeneous graphs and PLM-based node features
snapshot_download(repo_id="Cloudy1225/HTAG", repo_type="dataset", local_dir="./data", allow_patterns="*.pkl")

# Or just download raw texts
snapshot_download(repo_id="Cloudy1225/HTAG", repo_type="dataset", local_dir="./data", allow_patterns=["*.csv", "*.csv.zip"])

Dataset Format

The dataset includes heterogeneous graph edges, raw text, PLM-based features, labels, and years associated with text-attributed nodes. Raw text is provided in .csv or .csv.zip files, while the remaining data are stored in a dictionary object within a .pkl file. For example, by reading the tmdb.pkl file, the following dictionary can be obtained:

{'movie-actor': (array([   0,    0,    0, ..., 7504, 7504, 7504], dtype=int16),
  array([    0,     1,     2, ..., 11870,  1733, 11794], dtype=int16)),
 'movie-director': (array([   0,    0,    0, ..., 7503, 7503, 7504], dtype=int16),
  array([   0,    1,    2, ..., 3423,  966, 2890], dtype=int16)),
 'movie_labels': array([3, 1, 1, ..., 1, 1, 2], dtype=int8),
 'movie_feats': array([[ 0.00635284,  0.00649689,  0.01250827, ...,  0.06342042,
         -0.01747945,  0.0134356 ],
        [-0.14075027,  0.02825641,  0.02670695, ..., -0.12270895,
          0.08417314,  0.02486392],
        [ 0.00014208, -0.02286632,  0.00615967, ..., -0.03311544,
          0.04735276, -0.07458566],
        ...,
        [ 0.01835816,  0.07484645, -0.08099765, ..., -0.00150019,
          0.01669764,  0.00456595],
        [-0.00821487, -0.10434289,  0.01928608, ..., -0.06343049,
          0.05060194, -0.04229118],
        [-0.06465845,  0.13461556, -0.01640793, ..., -0.06274845,
          0.04002513, -0.00751513]], dtype=float32),
 'movie_years': array([2013, 1995, 1989, ..., 1939, 1941, 1965], dtype=int16)}

Dataset Statistics

	# Nodes	# Edges	# Classes	# Splits
TMDB	24,412	104,858	4	Train: 5,698
	Movie: 7,505	Movie-Actor: 86,517		Valid: 711
	Actor: 13,016	Movie-Director: 18,341		Test: 1,096
	Director: 3,891
CroVal	44386	164,981	6	Train: 980
	Question: 34153	Question-Question: 46,269		Valid: 1,242
	User: 8898	Question-User: 34,153		Test: 31,931
	Tag: 1335	Question-Tag: 84,559
ArXiv	231,111	2,075,692	40	Train: 47,084
	Paper: 81,634	Paper-Paper: 1,019,624		Valid: 18,170
	Author: 127,590	Paper-Author: 300,233		Test: 16,380
	FoS: 21,887	Paper-FoS: 755,835
Book	786,257	9,035,291	8	Train: 330,201
	Book	Book-Book: 7,614,902		Valid: 57,220
	Author	Book-Author: 825,905		Test: 207,063
	Publisher	Book-Publisher: 594,484
DBLP	1,989,010	29,830,033	9	Train: 508,464
	Paper: 964350	Paper-Paper: 16,679,526		Valid: 158,891
	Author: 958961	Paper-Author: 3,070,343		Test: 296,995
	FoS: 65699	Paper-FoS: 10,080,164
Patent	5,646,139	8,833,738	120	Train: 1,705,155
	Patent: 2,762,187	Patent-Inventor: 6,071,551		Valid: 374,275
	Inventor: 2,873,311	Patent-Examiner: 2,762,187		Test: 682,757
	Examiner: 10,641

Experiments

Datasets Setup

To set up the datasets, download them from HTAG · Datasets at Hugging Face into the ./data folder:

from huggingface_hub import snapshot_download

# Download heterogeneous graphs and PLM-based node features
snapshot_download(repo_id="Cloudy1225/HTAG", repo_type="dataset", local_dir="./data", allow_patterns="*.pkl")

Subsequently, the dataloader.py script will load the .pkl file, construct a dgl.DGLGraph object, and generate the train/validation/test splits.

Model Configuration

Model configurations are provided in the ./train.conf.yaml file.

Run Experiments

To reproduce the experiment results, execute the following commands:

python MLP.py

python GNN.py --model GCN/SAGE/GAT

python RGNN.py --model RGCN/RSAGE/RGAT/ieHGCN

	TMDB		CroVal		ArXiv		Book		DBLP		Patent
	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1
MLP	72.32±0.50	71.93±0.50	85.78±0.23	83.26±0.24	78.97±0.20	43.51±1.03	75.07±0.09	66.63±0.16	69.91±0.02	65.46±0.05	69.64±0.03	52.69±0.16
GCN	77.39±0.18	78.24±0.19	82.83±0.32	79.42±0.47	81.95±0.20	49.83±0.92	80.72±0.08	73.91±0.15	73.91±0.15	70.26±0.20	73.64±0.12	56.61±0.39
SAGE	79.29±0.57	80.06±0.59	85.26±0.37	82.84±0.45	83.89±0.14	52.18±0.54	81.43±0.07	74.83±0.11	75.62±0.07	72.43±0.09	75.58±0.14	59.54±0.79
GAT	79.69±0.27	80.40±0.23	83.37±0.36	80.56±0.36	83.84±0.41	50.24±1.82	80.96±0.07	74.19±0.16	75.62±0.20	72.49±0.28	74.74±0.12	56.30±0.49
RGCN	81.57±0.74	82.12±0.67	87.10±0.26	84.47±0.38	84.80±0.30	54.01±0.79	82.22±0.11	76.57±0.14	77.40±0.09	74.61±0.12	76.97±0.08	60.67±0.12
RSAGE	82.24±0.32	82.76±0.32	87.44±0.26	85.00±0.35	84.85±0.22	54.24±1.47	82.35±0.02	76.70±0.10	77.74±0.13	74.95±0.21	76.98±0.07	61.01±0.39
RGAT	81.97±0.16	82.40±0.18	87.43±0.18	84.93±0.22	84.69±0.25	54.62±1.49	82.45±0.04	76.75±0.06	77.81±0.29	74.98±0.33	77.22±0.08	61.36±0.18
ieHGCN	81.75±0.47	82.19±0.51	87.15±0.24	84.58±0.31	84.08±0.31	53.49±1.87	81.92±0.43	76.10±0.70	77.47±1.12	74.60±1.32	76.90±0.13	60.39±0.39