Pseudo-Code

Training:

Inference:

Experimental Results under Different Settings

Similar to Figure 4:

Error rate versus the number of random labelings during inference. We compare the embedding strategies for training, and all variants use the labeling with the maximum predicted probability for inference.

Similar to Figure 5:

Error rate versus the number of random labelings during both training and inference.

Dependenices

• tensorflow 1.13.2
• pytorch 1.6.0
• networkx 2.2