Improvement of Attention Mechanism Explainability in Prediction of Chemical Molecules’ Properties

Abstract

In this paper, the analysis of selected graph neural network operators is presented. The classic Graph Convolutional Network (GCN) was compared with methods containing trainable attention coefficients: Graph Attention Network (GAT) and Graph Transformer (GT). Moreover, which is an original contribution of this work, training of GT was modified with an additional loss function component enabling easier explainability of the produced model. The experiments were conducted using datasets with chemical molecules where both classification and regression tasks are considered. The results show that additional constraint not only does not make the results worse but, in some cases, it improves predictions.