Handling a graph dataset with different number of nodes/edges

[chinmay5]

I have a dataset with a different number of edges, ranging from 0 (no graph structure) all the way to 200 edges. Similarly, the graph nodes also range from 0 to 100. It is a binary label prediction task and I have the distribution of labels as follows:-

less than 80 edges has the distribution: {1 -> 121 and 0 -> 74}
more than 80 edges has the distribution: {1 -> 182 and 0 -> 53}

I think this indicates that the dataset has an inherent bias toward predicting label 1 when the number of edges is larger.

When I look at the situation with the labels would be:-
less than 40 noes : {1: ->182, 0 -> 53}
more than 40 noes {1 -> 60, 0-> 6}
Thus, a similar trend with the number of nodes.

As a result, I can observe that operations such as global_mean_pool and global_sum_pool for the readout work better than the global_max_pool. Each node represents an image and I use a pre-trained CNN to get the features.

My question is, what would be the preferred way to ensure that my model is not getting a bias based on the size of nodes and number of edges.

[rusty1s]

You can avoid such a bias by not adding graph structural features to the nodes (such as input degree), and do not make use of sum aggregation. For example, global_mean_pool(...) should work just fine since the model cannot really construct the number of nodes from it.

[chinmay5]

As of now, I have performed the k fold stratification by dividing the graphs into a group of 3. These are small, medium and large graphs. I think this is somewhat equivalent to the balanced sampler idea that we often use for imaging data. Do you think this makes sense?

[rusty1s]

Sure, if your goal is to find out whether your model is agnostic to graph sizes, this makes a lot of sense :)