Figure 1
From: Chemical property prediction under experimental biases
Main aim of our study. The left-hand figure shows the biased and unbiased distribution compared with the natural universe distribution of a chemical domain or sub-domain. The right-hand figure shows the proposed methods to train GNNs for chemical property prediction. Our aim was to apply bias cancelling techniques for GNNs to achieve significantly lower errors (i.e., MAE) when tested on a randomly sampled test dataset, whose distribution is similar to nature. G is the molecular graph, and \(y^k\) is the value of the k-th chemical property.
