SMRL

Spatial Meta-Learning-Based Representation for Unseen Geographic Entities

IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2026

Shengwen Li · Zhouzheng Xu · Renyao Chen · Jiarui Zhu · Yaqin Ye · Shunping Zhou · Hong Yao

---

TL;DR: SMRL learns transferable representations for geographic entities that were not observed during training. It builds spatially coherent meta-learning tasks, aggregates local relational and attribute context, and adapts the learned representation modules to an unseen geographic graph.

Highlights

• Unseen geographic entities. The target entities and their local graph context are introduced after representation learning on the known graph.
• Spatially aware task construction. Random-walk subgraphs form local support/query episodes instead of treating the geographic knowledge graph as an unstructured collection of triples.
• Relation–attribute initialization. Each entity combines incoming relation context with semantic attribute context before graph propagation.
• Transfer through meta-learning. SMRL optimizes representation modules across local tasks, then fine-tunes them for the unseen graph.

Method at a Glance

Stage	Purpose	Implementation
1. Spatial task construction	sample local training episodes from the known geographic graph	subgraph.py, datasets.py
2. Local representation	initialize and propagate entity context from relations and attributes	new_ent_init_model.py, rgcn_model.py
3. Meta-train and adapt	learn transferable parameters and fine-tune on unseen entities	meta_trainer.py, post_trainer.py

For an entity with incoming relation embeddings R and attribute embeddings T, the initializer uses:

h = mean(R) + mean(T)

Attribute facts include non-entity-object triples and semantic metadata relations such as rdf:type, rdfs:label, RegionId, and worldkg.org/schema/*. Attribute keys are indexed as (predicate, object) pairs.

Installation

The reference environment uses Python with:

torch==1.7.1
dgl==0.6.1
lmdb>=0.99
numpy>=1.19.0
tensorboard>=2.4.0
tqdm>=4.60.0

Install the listed dependencies, selecting PyTorch and DGL builds that match your CUDA runtime:

pip install -r requirements.txt

Data Preparation

Place the known and unseen graph folders under data/:

data/
├── region_v6/
│   ├── train.txt
│   ├── valid.txt
│   └── test.txt
└── region_6_ind/
    ├── train.txt
    ├── valid.txt
    └── test.txt

Each file contains one triple per line:

<head>*<relation>*<tail>

The parser accepts both * and ^ separators. See data/sample_region_v6 and data/sample_region_6_ind for format examples. The first run creates cached pickle files and LMDB subgraph databases under data/.

Training

1. Meta-train on the known graph

python main.py \
  --data_name region_v6 \
  --ind_data_name region_6_ind \
  --name region_v6_ComplEx_smrl \
  --step meta_train \
  --kge ComplEx \
  --gpu cuda:0 \
  --use_attr true

2. Fine-tune on the unseen graph

python main.py \
  --data_name region_v6 \
  --ind_data_name region_6_ind \
  --name region_v6_ComplEx_smrl_finetune \
  --metatrain_state ./state/region_v6_ComplEx_smrl/region_v6_ComplEx_smrl.best \
  --step fine_tune \
  --kge ComplEx \
  --gpu cuda:0 \
  --use_attr true

Equivalent launchers are provided in script/metatrain.sh and script/finetune.sh. Outputs are written to state/, log/, and tb_log/.

Configuration Notes

• --ind_data_name explicitly binds a known graph to its unseen graph. If omitted, region_v6 is matched with region_6_ind when that folder exists.
• --num_attr is detected during preprocessing. Set it manually only when loading a checkpoint with a fixed attribute vocabulary.
• --attr_weight controls the contribution of aggregated attribute embeddings.
• --force_preprocess rebuilds the cached graph and attribute indices.

Citation

If SMRL is useful in your research, please cite:

@article{li2026smrl,
  author  = {Li, Shengwen and Xu, Zhouzheng and Chen, Renyao and Zhu, Jiarui and Ye, Yaqin and Zhou, Shunping and Yao, Hong},
  title   = {Spatial Meta-Learning-Based Representation for Unseen Geographic Entities},
  journal = {IEEE Transactions on Neural Networks and Learning Systems},
  year    = {2026},
  pages   = {1--13},
  doi     = {10.1109/TNNLS.2026.3679789}
}