Adversarial Permutation Guided Node Representations for Link Prediction

After observing a snapshot of social network, link prediction (LP) algorithm identifies node pairs between which new edges will likely materialize in future. Most LP algorithms estimate score for currently non-neighboring pairs, and rank them by this score. Recent systems compute comparing dense, low dimensional vector representations nodes. Graph neural networks (GNNs), particular graph convolutional (GCNs), are popular examples. For two nodes to be meaningfully compared, their embeddings should indifferent reordering neighbors. GNNs typically use simple, symmetric set aggregators ensure property, but design decision has been shown produce with limited expressive power. Sequence encoders more expressive, permutation sensitive design. efforts overcome dilemma turn out unsatisfactory tasks. In response, we propose PermGNN, aggregates neighbor features using recurrent, order-sensitive aggregator directly minimizes an loss while it is `attacked' adversarial generator permutations. PermGNN superior power compared earlier GNNs. Next, devise optimization framework map PermGNN's suitable locality-sensitive hash, speeds up reporting the top-K most task. Our experiments on diverse datasets show that outperforms several state-of-the-art predictors significant margin, can predict fast.