Modeling and Mining Graphs with Feature-Rich Nodes

Friday, May 11, 2018

11.30 a.m.

ISI seminar room 1st floor

Dr. Corrado Monti Università degli Studi di Milano

Real-world complex networks describe connections between objects; in reality, those objects are often endowed with some features. How does their presence interplay with the network link structure? Many previous works considered homophily as the only possible transmission mechanism translating node features into links. We want instead to be able to represent a wide range of scenarios — not only homophily and heterophily – but still be able to deal with very large networks (10^7 nodes). For this reason, we expand on previous models, where interactions between pairs of features can foster or discourage links. To estimate those latent interactions, we propose two solutions: the first one based on Naive Bayes; the second one on online machine learning. We use this model to devise learning algorithms for graph mining problems – such as link prediction, anomaly detection, and discovering missing features.
Finally, we will present some experimental results on large collaboration networks and on semantic networks.