Graph convolutional neural networks with node transition probability-based message passing and DropNode regularization

Graph convolutional neural networks (GCNNs) have received much attention recently, owing to their capability in handling graph-structured data. Among the existing GCNNs, many methods can be viewed as instances of a message passing motif; features nodes are passed around neighbors, aggregated and transformed produce better nodes’ representations. Nevertheless, these seldom use node transition probabilities, measure that has been found useful exploring graphs. Furthermore, when probabilities used, direction is often improperly considered feature aggregation step, resulting an inefficient weighting scheme. In addition, although great number GCNN models with increasing level complexity introduced, GCNNs suffer from over-fitting being trained on small Another issue over-smoothing, which tends make representations indistinguishable. This work presents new method improve process based by properly considering direction, leading scheme compared counterpart. Moreover, we propose novel regularization termed DropNode address over-smoothing issues simultaneously. randomly discards part graph, thus it creates multiple deformed versions data augmentation effect. Additionally, lessens connectivity mitigating effect deep GCNNs. Extensive experiments eight benchmark datasets for graph classification tasks demonstrate effectiveness proposed comparison state art.