Interpretable Signed Link Prediction With Signed Infomax Hyperbolic Graph

Signed link prediction in social networks aims to reveal the underlying relationships (i.e., links) among users nodes) given their existing positive and negative interactions observed. Most of prior efforts are devoted learning node embeddings with graph neural (GNNs), which preserve signed network topology by message-passing along edges facilitate downstream task. Nevertheless, graph-based approaches could hardly provide human-intelligible explanations for following three questions: (1) neighbors aggregate, (2) path propagate along, (3) theory follow process. To answer aforementioned questions, this paper, we investigate how reconcile balance xmlns:xlink="http://www.w3.org/1999/xlink">status rules information develop a unified framework, termed as Infomax Hyperbolic Graph ( SIHG ). By maximizing mutual between edge polarities embeddings, one can identify most representative neighboring nodes that support inference sign. Different from GNNs only group features friends subspace, proposed SIHG incorporates attention module, is also capable pushing hostile far away each other geometry antagonism. The polarity learned maps, turn, provides interpretations theories used aggregation. In order model high-order user relations complex hierarchies, projected measured hyperbolic space lower distortion. Extensive experiments on four benchmarks demonstrate framework significantly outperforms state-of-the-arts prediction.