I'm trying to compare 2 models GraphConv and GCNConv for my project. According to the tutorial provided by Pytorch GraphConv preserves central node information by omitting neighbourhood normalisation. The tutorial shows that GraphConv has improved the test accuracy to 82%, compared to 76%, in the case of the GCNConv model. However, in my case GraphConv increased the training accuracy from 89% (by GCNConv) to 93% but no change in test accuracy i.e. 88% (similar in GraphConv and GCNConv). I'm keeping the rest of the parameters the same, so what could be the reason(s)? Is there any way to increase the test accuracy? Thank you, below is the link to the Pytorch tutorial I mentioned here. I tried to compare GraphConv and GCNConv, GraphConv is supposed to give me better test accuracy results, but it doesn't. Continue reading...
