The different results of unsupervised training on TU dataset.

[ZJUDataIntelligence]

When I reproduced your code, I ran the un_exp part of the code and got a result, which is the first line in the table below. This is the result I got when I found that your masking option for node features was False. Then I changed the node feature masking option to True and got the second line of results. The third line is the result in your paper. It’s strange why the variance of the first two lines of results is so large, and why the line without node masking has higher results than others.

	MUTAG	PROTEINS	REDDIT-B	DD	IMDB-B	NCI1	COLLAB	REDDIT-M-5K
no-mask	90.99±7.80	76.37±3.26	90.30±1.99	78.53±3.58	73.90±3.33	82.34±1.16	72.18±2.07	56.69±1.63
mask	89.88±5.57	76.83±3.36	88.55±2.47	77.42±3.18	72.70±4.36	82.63±1.72	70.64±1.59	56.41±2.28
paper	88.64±1.08	75.80±0.36	88.58±1.49	77.57±0.60	73.30±0.40	82.00±0.29	70.12±0.68	56.75±0.18