mgalkin/ultra_3g

Description

ULTRA is a foundation model for knowledge graph (KG) reasoning. A single pre-trained ULTRA model performs link prediction tasks on any multi-relational graph with any entity / relation vocabulary. Performance-wise averaged on 50+ KGs, a single pre-trained ULTRA model is better in the 0-shot inference mode than many SOTA models trained specifically on each graph. Following the pretrain-finetune paradigm of foundation models, you can run a pre-trained ULTRA checkpoint immediately in the zero-shot manner on any graph as well as use more fine-tuning.

ULTRA provides unified, learnable, transferable representations for any KG. Under the hood, ULTRA employs graph neural networks and modified versions of NBFNet. ULTRA does not learn any entity and relation embeddings specific to a downstream graph but instead obtains relative relation representations based on interactions between relations.

arxiv: https://arxiv.org/abs/2310.04562
GitHub: https://github.com/DeepGraphLearning/ULTRA

Checkpoints

Here on HuggingFace, we provide 3 pre-trained ULTRA checkpoints (all ~169k params) varying by the amount of pre-training data.

Model	Training KGs
ultra_3g	3 graphs
ultra_4g	4 graphs
ultra_50g	50 graphs

• ultra_3g and ultra_4g are the PyG models reported in the github repo;
• ultra_50g is a new ULTRA checkpoint pre-trained on 50 different KGs (transductive and inductive) for 1M steps to maximize the performance on any unseen downstream KG.

⚡️ Your Superpowers

ULTRA performs link prediction (KG completion aka reasoning): given a query (head, relation, ?), it ranks all nodes in the graph as potential tails.

1. Install the dependencies as listed in the Installation instructions on the GitHub repo.
2. Clone this model repo to find the UltraForKnowledgeGraphReasoning class in modeling.py and load the checkpoint (all the necessary model code is in this model repo as well).

• Run zero-shot inference on any graph:

from modeling import UltraLinkPrediction
from ultra.datasets import CoDExSmall
from ultra.eval import test

model = UltraForKnowledgeGraphReasoning.from_pretrained("mgalkin/ultra_3g")
dataset = CoDExSmall(root="./datasets/")
test(model, mode="test", dataset=dataset, gpus=None)
# Expected results for ULTRA 3g
# mrr:      0.472
# hits@10:  0.668

Or with AutoModel:

from transformers import AutoModel
from ultra.datasets import CoDExSmall
from ultra.eval import test

model = AutoModel.from_pretrained("mgalkin/ultra_3g", trust_remote_code=True)
dataset = CoDExSmall(root="./datasets/")
test(model, mode="test", dataset=dataset, gpus=None)
# Expected results for ULTRA 3g
# mrr:      0.472
# hits@10:  0.668

• You can also fine-tune ULTRA on each graph, please refer to the github repo for more details on training / fine-tuning
• The model code contains 57 different KGs, please refer to the github repo for more details on what's available.

Performance

Averaged zero-shot performance of ultra-3g and ultra-4g

Model	Inductive (e) (18 graphs)		Inductive (e,r) (23 graphs)		Transductive (16 graphs)
	Avg MRR	Avg Hits@10	Avg MRR	Avg Hits@10	Avg MRR	Avg Hits@10
ULTRA (3g) PyG	0.420	0.562	0.344	0.511	0.329	0.479
ULTRA (4g) PyG	0.444	0.588	0.344	0.513	WIP	WIP
ULTRA (50g) PyG (pre-trained on 50 KGs)	0.444	0.580	0.395	0.554	0.389	0.549

Fine-tuning ULTRA on specific graphs brings, on average, further 10% relative performance boost both in MRR and Hits@10. See the paper for more comparisons.

ULTRA 50g Performance

ULTRA 50g was pre-trained on 50 graphs, so we can't really apply the zero-shot evaluation protocol to the graphs. However, we can compare with Supervised SOTA models trained from scratch on each dataset:

Model	Avg MRR, Transductive graphs (16)	Avg Hits@10, Transductive graphs (16)
Supervised SOTA models	0.371	0.511
ULTRA 50g (single model)	0.389	0.549

That is, instead of training a big KG embedding model on your graph, you might want to consider running ULTRA (any of the checkpoints) as its performance might already be higher 🚀

Useful links

Please report the issues in the official GitHub repo of ULTRA