This is a list of sampling strategies implemented in this extension.
For the --sampler argument use the name in parentheses.
N randomly sampled nodes are selected and their corresponding adjacency information and node features are aggregated.
Non-uniform sampling, probability of a node being sampled is relative to its PageRank weight.
Non-uniform sampling, the higher a nodes degree the higher the probability to be sampled; leads to very dense samples.
Create a set S of the top k nodes by ranking the nodes according to their degree. Rank the neighbors of a random node from S. In order to create the new seed set S for the following iteration, the top k nodes of these iterations are added to the sub-graph.
Same as Rank Degree but in contrast we maintain a list of candidate nodes that is composed of all the neighbors of nodes that have been found but not yet sampled. Experiments show that this method of sampling retains a better balance between the depth and breadth of the sampled sub-graph.
Generates n source-destination pairs that do not share an edge. It is possible that in small or very dense graphs less than n pairs are generated, because less than n pairwise-unconnected nodes in the graph exist.
Edges are uniformly sampled instead of the nodes themselves.
Select a starting node and select an outgoing edge from the starting node at random. Add the connected node to the sample. Repeat this procedure with a new node selected at random. This method has less bias towards high degree nodes than Random Edge.
Do one step of Random Node Edge with probability p (set to 0.8 in the paper) and a step of Random Edge otherwise.
Uniformly sample random nodes and add a fixed number k of its neighbor. This keeps the computational expense constant for every batch.
Select a random node and also add all its (outgoing) neighbors; might explode in dense graphs.
Select a random node and 'walk' along randomly selected outgoing edges. With probability of 0.15 (often used in literature) go back to first node and restart walk. Can get stuck, but possible to use iteration counter at which to break and select a new start node.
Same as Random Walk, but prevents getting stuck by jumping to a new random node with probability of 0.15 instead of only going back to start.
Select a start node and start adding (burning) edges with corresponding nodes. When an edge and its corresponding other node is burned, this node can also add its edges.
By keeping track of a list of m nodes that represent the graph, M-dimensional dependent random walks can sample the data. Choose a node at random from the list, take a random walk, and swap it out for the newly chosen node. The edge which was chosen by the random walk is added to sample set.
The algorithm makes an effort to select nodes from every community in a graph. By maximizing an expansion factor, the algorithm aims to construct the sample in a greedy manner.