A signed directed network is a network where nodes have signed and directed connections that reflect differenttypes of relationships in various domains and that can help users understand network dynamics and patterns. Nodeembedding is used to map nodes to vectors while preserving their properties by placing similar nodes close in vectorspace. Here, the similarity criterion can be proximity-based (closeness of nodes in networks) or structure-based(similarity of connection patterns). Despite their high expressive capability and potential in practical applications,structure-based embedding methods for signed directed networks still need further development. To address thisgap, we propose SDs2vec, a novel structural embedding method for signed directed networks, and we implementmultiple optimization strategies to very efficiently reduce computational costs. We conducted intensive experimentson toy and real networks to prove the method’s high interpretability and performance for downstream tasks that varyfrom edges’ sign and direction prediction to nodes’ degree prediction. The results show that the proposed methodoutperforms state-of-the-art methods in most cases, demonstrating its high capability in practical applications such aslink prediction and node classification.
Liu et al. (Sat,) studied this question.