This is a Pytorch implementation for our submitted AI OPEN journal paper: Structure-Enhanced Meta-Learning For Few-Shot Graph Classification[arXiv]（http://arxiv.org/abs/2103.03547）

Contributors: Shunyu Jiang, Fuli Feng, Weijian Chen, Xiangnan He

We used the following Python packages for core development. We tested on Python 3.7.

pytorch                   1.0.1
rdkit                     2019.03.1.0
numpy                     
json
pandas

+ --lr:               learning rate
+ --type:             local-structure or global-structure
+ --norm_type：       centering and scaling in paper
+ --attention_type:   five attention models in paper
+ --num_ways:         classes number in meta-task in paper
+ --spt_shots:        support-set number in meta-task in paper
+ --qry_shots:        query-set number in meta-task in paper

We conduct experiments on the multi-class Chembl dataset and public dataset TRIANGLES. Detailed information about these two datasets illustrated in paper. Dataset ChemBL was too large, so we divided it into 273 subfiles according to classes. Finally, we compressed and uploaded the ChemBL and TRIANGLES datasets.

We provide solutions of local structure and global structure respectively, which contain five different models of attention mechanism.

nohup python main_local.py --type=local --attention_type=transformer --dataset=TRIANGLES --num_ways=3 > TRIANGLES_local_log.out &

or

nohup python main_global.py --type=gloabl --attention_type=self-attention --dataset=TRIANGLES --num_ways=3 > TRIANGLES_global_log.out &