Model Graph Inductive Learning (MGIL) for Knowledge Graph Completion

This repository contains the official implementation of the paper:

---

Overview

Link prediction in knowledge graphs relies heavily on high-quality embeddings. However, most existing approaches focus only on local neighborhood aggregation and ignore the global structure of the graph. To address this limitation, we propose MGIL (Model Graph Inductive Learning), a novel framework that:

• Constructs a model graph from the original knowledge graph
• Captures global structural patterns
• Generates high-quality initial embeddings for entities

---

Key Idea

MGIL builds two type model graphs using two strategies:

1. Relation-based Clustering

Entities are grouped based on the similarity of their:

• Incoming relations
• Outgoing relations

2. Type-based Clustering

Entities are grouped based on their semantic types:

• Example: drugs, proteins, diseases

Edge in Model Graph

In the model graph, nodes represent groups of entities that share identical relational feature vectors. An undirected edge between two nodes (U_j) and (U_k) is created if there exists at least one triple ((v_p, r, v_q)) in the knowledge graph such that (v_p \in U_j) and (v_q \in U_k), where (r \in \mathcal{R}).

This means that if any entity in group (U_j) is connected to any entity in group (U_k) through any observed relation in the original knowledge graph, an edge is added between the corresponding model graph nodes. The edge captures aggregated interactions between entity groups rather than individual entity-level connections.

A Graph Neural Network (GNN) is then applied to the model graph to learn embeddings, which are transferred to the original graph.

---

Framework Pipeline

1. Construct model graph (relation-based or type-based)
2. Apply GNN on the model graph
3. Generate global-aware embeddings
4. Initialize original KG embeddings
5. Perform link prediction

Framework Overview

Model Graph Construction

Relation-based Model Graph

Entity-type Model Graph

Inductive Datasets

We evaluate MGIL on several widely-used and recently proposed inductive knowledge graph completion benchmarks:

• FB15k-237

• WN18RR

• NELL-995

• Shomer Benchmarks:

  • CoDEx-M_E
  • WN18RR_E
  • HetioNet_E

---

Benchmark Summary

Benchmark	Directory	Datasets	model_graph_type
Grail	dataset	nell_v1 ~ nell_v4	relation_base
Grail	dataset	fb237_v1 ~ fb237_v4	relation_base
Grail	dataset	wn18rr_v1 ~ wn18rr_v4	relation_base
Shomer	dataset/new_data	codex_m_E	relation_base
Shomer	dataset/new_data	wn18rr_E	relation_base
Shomer	dataset/new_data	hetionet_E	relation_base or entity_base

Note: Shomer datasets support two inductive inference settings:

• inference_1 — evaluates on test split 1 (test_0_graph.txt / test_0_samples.txt)
• inference_2 — evaluates on test split 2 (test_1_graph.txt / test_1_samples.txt)

---

🚀 How to Run

We provide a simple command-line interface to train and evaluate the MGIL framework.

Quick Reference

Grail Benchmark (--benchmark dataset)

Dataset	Command
nell_v1	python main.py --step meta_train --data_name nell_v1 --benchmark dataset --model_graph_type relation_base
nell_v2	python main.py --step meta_train --data_name nell_v2 --benchmark dataset --model_graph_type relation_base
nell_v3	python main.py --step meta_train --data_name nell_v3 --benchmark dataset --model_graph_type relation_base
nell_v4	python main.py --step meta_train --data_name nell_v4 --benchmark dataset --model_graph_type relation_base
fb237_v1	python main.py --step meta_train --data_name fb237_v1 --benchmark dataset --model_graph_type relation_base
fb237_v2	python main.py --step meta_train --data_name fb237_v2 --benchmark dataset --model_graph_type relation_base
fb237_v3	python main.py --step meta_train --data_name fb237_v3 --benchmark dataset --model_graph_type relation_base
fb237_v4	python main.py --step meta_train --data_name fb237_v4 --benchmark dataset --model_graph_type relation_base
wn18rr_v1	python main.py --step meta_train --data_name wn18rr_v1 --benchmark dataset --model_graph_type relation_base
wn18rr_v2	python main.py --step meta_train --data_name wn18rr_v2 --benchmark dataset --model_graph_type relation_base
wn18rr_v3	python main.py --step meta_train --data_name wn18rr_v3 --benchmark dataset --model_graph_type relation_base
wn18rr_v4	python main.py --step meta_train --data_name wn18rr_v4 --benchmark dataset --model_graph_type relation_base

Shomer Benchmark (--benchmark dataset/new_data)

Dataset	model_graph_type	Command
codex_m_E	relation_base	python main.py --step meta_train --data_name codex_m_E --benchmark dataset/new_data --model_graph_type relation_base
wn18rr_E	relation_base	python main.py --step meta_train --data_name wn18rr_E --benchmark dataset/new_data --model_graph_type relation_base
hetionet_E	relation_base	python main.py --step meta_train --data_name hetionet_E --benchmark dataset/new_data --model_graph_type relation_base
hetionet_E	entity_base	python main.py --step meta_train --data_name hetionet_E --benchmark dataset/new_data --model_graph_type entity_base

Fine-tuning

python main.py \
  --step fine_tune \
  --data_name nell_v1 \
  --benchmark dataset \
  --model_graph_type relation_base \
  --metatrain_state ./state/nell_v1/nell_v1.best

### 🔹 1. Meta-Training

Train the model on base datasets:

```bash
python main.py \
  --step meta_train \
  --data_name codex_m_E \
  --model_graph_type relation_base \
  --kge TransE \
  --num_layers 3 \
  --emb_dim 32

---

Key Arguments

Argument	Type	Default	Choices	Description
--data_name	str	nell_v1	—	Name of the dataset to use
--benchmark	str	dataset	dataset, dataset/new_data	Benchmark type: dataset for Grail, dataset/new_data for Shomer
--model_graph_type	str	relation_base	relation_base, entity_base	Model graph construction strategy
--step	str	meta_train	meta_train, fine_tune	Training mode
--test_type	str	inference_1	inference_1, inference_2	Inductive test split (Shomer datasets only)
--kge	str	TransE	TransE, DistMult, ComplEx, RotatE	Knowledge graph embedding model
--emb_dim	int	32	—	Embedding dimension
--num_layers	int	3	—	Number of R-GCN layers
--num_bases	int	4	—	Number of bases for R-GCN weight decomposition
--batch_size	int	64	—	Batch size for training
--lr	float	0.01	—	Learning rate
--gamma	float	10.0	—	Margin parameter for KGE loss
--adv_temp	float	1.0	—	Temperature for adversarial negative sampling
--num_neg	int	32	—	Number of negative samples per positive
--train_num_epoch	int	3	—	Number of meta-training epochs
--posttrain_num_epoch	int	50	—	Number of post-training epochs
--seed	int	1234	—	Random seed for reproducibility
--gpu	str	cuda:0	—	GPU device identifier
--metatrain_state	str	./state/fb237_v1_transe/fb237_v1_transe.best	—	Path to pre-trained state file (required for --step fine_tune)

---

Subgraph Parameters

Argument	Type	Default	Description
--num_train_subgraph	int	10000	Number of training subgraphs
--num_valid_subgraph	int	200	Number of validation subgraphs
--num_sample_for_estimate_size	int	50	Number of samples for size estimation
--rw_0	int	10	Random walk parameter 0
--rw_1	int	10	Random walk parameter 1
--rw_2	int	5	Random walk parameter 2
--num_sample_cand	int	5	Number of sample candidates

---

Model Graph Parameters

Argument	Type	Default	Description
--is_weighted_model_graph	bool	False	Use weighted edges in model graph
--is_directed_model_graph	bool	False	Use directed edges in model graph
--indtest_eval_bs	int	512	Batch size for inductive test evaluation
--metatrain_check_per_step	int	625	Checkpoint interval during meta-training
--posttrain_check_per_epoch	int	625	Checkpoint interval during post-training

Authors

• Mohammad Esmail Khani
• Hossein Hajiabolhassan
• Ali Taherkhani

Citation

• Mohommad Esmaeil Khani, Mahdieh Hasheminejad, Ali Taherkhani, and Hossein Hajiabolhassan, Model Graph Inductive Learning for Knowledge Graph Completion, arXiv:2606.16509, 2026.