dgraphxinye/2022_finvcup_baseline

第七届信也科技杯baseline

这是第七届信也科技杯-欺诈用户风险识别的baseline。
请在比赛网站上下载"初赛数据集.zip"文件，将zip文件中的"phase1_gdata.npz"放到路径'./xydata/raw'中。
baseline代码中对"phase1_gdata.npz"的train_mask，随机按照6/4的比例将其划分为train/valid dataset。

Environments

Implementing environment:

• python = 3.7.6

• numpy = 1.21.2

• pytorch = 1.6.0

• torch_geometric = 1.7.2

• torch_scatter = 2.0.8

• torch_sparse = 0.6.9

• GPU: Tesla V100 32G

Training

• MLP

python train.py --model mlp  --epochs 200 --device 0
python inference.py --model mlp --device 0

• GCN

python train.py --model gcn  --epochs 200 --device 0
python inference.py --model gcn --device 0

• GraphSAGE

python train.py --model sage  --epochs 200 --device 0
python inference.py --model sage --device 0

• GraphSAGE (NeighborSampler)

python train_mini_batch.py --model sage_neighsampler --epochs 200 --device 0
python inference_mini_batch.py --model sage_neighsampler --device 0

• GAT (NeighborSampler)

python train_mini_batch.py --model gat_neighsampler --epochs 200 --device 0
python inference_mini_batch.py --model gat_neighsampler --device 0

• GATv2 (NeighborSampler)

python train_mini_batch.py --model gatv2_neighsampler --epochs 200 --device 0
python inference_mini_batch.py --model gatv2_neighsampler --device 0

Results:

在以上的依赖环境中，baseline中几个模型效果如下：

Methods	Train AUC	Valid AUC	Test AUC
MLP	0.7305	0.7328	0.7283
GCN	0.7272	0.7336	0.7333
GraphSAGE	0.7799	0.7798	0.7727
GraphSAGE (NeighborSampler)	0.7916	0.7875	0.7810
GAT (NeighborSampler)	0.7462	0.7411	0.7329
GATv2 (NeighborSampler)	0.7818	0.7804	0.7733

2022_finvcup_baseline's Issues

assert self.lin_edge is not None

跑 Gat的minibatch的时候报错

 File "//2022_finvcup_baseline/models/gat_neighsampler.py", line 158, in forward
    x = self.convs[i]((x, x_target), edge_index)
  File "/miniconda3/envs/torch17/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/miniconda3/envs/torch17/lib/python3.7/site-packages/torch_geometric/nn/conv/gatv2_conv.py", line 236, in forward
    size=None)
  File "/miniconda3/envs/torch17/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 317, in propagate
    out = self.message(**msg_kwargs)
  File "/miniconda3/envs/torch17/lib/python3.7/site-packages/torch_geometric/nn/conv/gatv2_conv.py", line 266, in message
    assert self.lin_edge is not None

但是跑 SAG的minibatch是可以正常训练的,这是我用的版本不对吗?

torch-geometric         2.0.4
torch-scatter           2.0.6
torch-sparse            0.6.9