Structural Node Embedding in Signed Social Networks: Finding Online Misbehavior at Multiple Scales

Entities in networks may interact positively as well as negatively with each other, which may be modeled by a signed network containing both positive and negative edges between nodes. Understanding how entities behave and not just with whom they interact positively or negatively leads us to the new problem of structural role mining in signed networks. We solve this problem by developing structural node embedding methods that build on sociological theory and technical advances developed specifically for signed networks. With our methods, we can not only perform node-level role analysis, but also solve another new problem of characterizing entire signed networks to make network-level predictions. We motivate our work with an application to social media analysis, where we show that our methods are more insightful and effective at detecting user-level and session-level malicious online behavior from the network structure than previous approaches based on feature engineering.