Understanding the message passing in graph neural networks via power iteration clustering

The mechanism of message passing in graph neural networks (GNNs) is still mysterious. Apart from convolutional networks, no theoretical origin for GNNs has been proposed. To our surprise, can be best understood terms power iteration. By fully or partly removing activation functions and layer weights GNNs, we propose subspace iteration clustering (SPIC) models that iteratively learn with only one aggregator. Experiments show extend enhance their capability to process random featured networks. Moreover, demonstrate the redundancy some state-of-the-art design define a lower limit model evaluation by aggregator passing. Our findings push boundaries understanding