AutoGF: Runtime Graph Filter Tuning for Community Node Ranking (preprint of accepted manuscript)

A recurring graph analysis task is to rank nodes based on their relevance to overlapping communities of shared metadata attributes (e.g. the interests of social network users). To achieve this, approaches often start with a few example community members and employ graph filters that rank nodes based on their structural proximity to the examples. Choosing between well-known filters typically involves experiments on existing graphs, but their efficacy is known to depend on the structural relations between community members. Therefore, we argue that employed filters should be determined not during algorithm design but at runtime, upon receiving specific graphs and example nodes to process. To do this, we split example nodes into training and validation sets and either perform supervised selection between well-known filters, or account for granular graph dynamics by tuning parameters of the generalized graph filter form with a novel optimization algorithm. Experiments on 27 community node ranking tasks across three real-world networks of various sizes reveal that runtime algorithm selection selects near-best AUC and NDCG among a list of 8 popular alternatives, and that parameter tuning yields similar or improved results in all cases.