PointPattern

PointPattern is a graph classification dataset constructed by simple point patterns from statistical mechanics. The authors simulated three point patterns in 2D: hard disks in equilibrium (HD), Poisson point process, and random sequential adsorption (RSA) of disks. The HD and Poisson distributions can be seen as simple models that describe the microstructures of liquids and gases while the RSA is a nonequilibrium stochastic process that introduces new particles one by one subject to nonoverlapping conditions.

These systems are well known to be structurally different, while being easy to simulate, thus provides a solid and controllable classification task. For each point pattern, the particles are treated as nodes, and edges are subsequently drawn according to whether two particles are within a threshold distance.

Source: Path Integral Based Convolution and Pooling for Graph Neural Networks Image Source: Path Integral Based Convolution and Pooling for Graph Neural Networks