Genius: Subteam Replacement with Clustering-based Graph Neural Networks

The state of the art for subteam replacement, based on random walk graph kernels, encounter the following limitations: (1) ineffective in capturing fine-grained node feature correlations, (2) inefficient without proper pruning mechanisms, and (3) limited applicability to single-member or equal-sized subteam replacements. In this paper, we address these limitations by proposing Genius, a clustering-based graph neural network (GNN) framework that (1) captures team social network knowledge for subteam replacement by deploying team-level attention GNNs (TAGs) and self-supervised positive team contrasting training scheme, (2) generates unsupervised team social network member clusters to prune candidates for fast computation, and (3) incorporates a subteam recommender that selects new subteams of flexible sizes. We demonstrate the efficacy of the proposed method in terms of (1) effectiveness: being able to select better subteam members that significantly increase the similarity between the new and original teams, and (2) efficiency: achieving more than 600× speed-up in average running time.