UniHG: A Large-scale Universal Heterogeneous Graph Dataset and Benchmark

Abstract

We present UniHG, the largest universal heterogeneous graph dataset containing 77.31M nodes and 564M edges with 2,082 relation types. Addressing three fundamental challenges in HG research—benchmark deficiency, semantic misalignment, and propagation degradation—we propose:

A semantic alignment strategy projecting multi-attribute features into unified embedding space
Heterogeneous Graph Decoupling (HGD) framework with Anisotropic Feature Propagation (AFP)

Extensive experiments show 28.93% accuracy improvement over SOTA methods and 11.71% NDCG@20 boost in downstream tasks. Dataset & Code

Metric	UniHG	MAG240M	Amazon Review
Nodes	77.31M	244.16M	102.70M
Edges	564M	1.73B	571.54M
Node Types	1	3	33
Edge Types	2,082	3	-
Labels	74,666	153	-

Experimental Results

Node Classification Performance

Method	UniHG-1M			UniHG-10M			UniHG-Full
Method	Accuracy	Recall	F1	Accuracy	Recall	F1	Accuracy	Recall	F1
HGD (Ours)	75.41	75.95	82.64	89.03	90.11	93.05	93.16	93.83	96.09
GCN	23.68	25.26	21.89	22.86	25.66	23.19	26.72	24.19	23.26
GAT	29.12	34.02	27.76	33.73	37.48	29.88	31.02	35.07	26.92
HGT	51.28	55.30	51.55	52.09	56.52	53.03	55.36	61.91	56.47
MTMP	47.52	47.48	60.79	59.95	61.33	72.81	65.67	66.15	67.74
SGC	42.56	42.11	55.12	56.32	57.78	67.95	62.15	63.88	73.21
SIGN	56.73	55.54	69.41	73.58	84.17	80.30	69.04	70.49	81.48
GAMLP	44.55	43.74	56.92	59.47	61.32	70.02	64.23	66.05	74.34

All values in percentage (%)

Ablation Study with AFP Module

Method	UniHG-1M			UniHG-10M			UniHG-Full
	Acc	Rec	F1	Acc	Rec	F1	Acc	Rec	F1
SGC	42.56	42.11	55.12	56.32	57.78	67.95	62.15	63.88	73.21
SGC+AFP	44.18	42.75	57.65	64.25	65.21	76.67	69.84	71.40	81.45
SIGN	56.73	55.54	69.41	73.58	84.17	80.30	69.04	70.49	81.48
SIGN+AFP	66.69	65.27	77.16	77.52	80.18	81.38	75.45	76.42	85.71
GAMLP	44.55	43.74	56.92	59.47	61.32	70.02	64.23	66.05	74.34
GAMLP+AFP	47.24	46.80	60.85	60.99	62.70	73.38	72.89	74.95	83.51

Cross-domain Knowledge Transfer

Method	Amazon-Book			Yelp2018			Citeulike-a
	Prec@20	Rec@20	NDCG@20	Prec@20	Rec@20	NDCG@20	Prec@20	Rec@20	NDCG@20
LightGCN	0.01716	0.06191	0.04106	0.00433	0.01123	0.00849	0.02329	0.07188	0.04374
LightGCN+UniHG	+2.797%	+1.712%	+0.803%	+6.467%	+7.925%	+5.535%	+3.521%	+3.116%	+1.935%
NGCF	0.01540	0.06278	0.04142	0.00318	0.00596	0.00533	0.03992	0.11455	0.09264
NGCF+UniHG	+14.675%	+17.776%	+27.764%	+6.918%	+31.543%	+9.005%	+0.601%	+1.422%	+1.759%

UniHG: A Large-scale Universal Heterogeneous Graph Dataset and Benchmark for Representation Learning and Cross-Domain Transferring

Anonymous Authors

Abstract

Dataset Statistics

Experimental Results

Node Classification Performance

Ablation Study with AFP Module

Cross-domain Knowledge Transfer

Conclusion

References