- Find a label component as all connected nodes of the same label - i.e, a maximal set of data points that are relevant and related to each other - randomly select 100 nodes as seeds from this set of data points