A0703
Title: Likelihood-based methods for partially observed networks
Authors: Udbhav Dalavai - Birkbeck, University of London (United Kingdom) [presenting]
Swati Chandna - Birkbeck, University of London (United Kingdom)
Abstract: Partially observed networks are common in many real-world applications due to data limitations, privacy concerns, or experimental limitations. In such settings, an appropriate statistical methodology for predicting missing links is crucial for enabling complete and meaningful analysis. The task of link prediction is studied under the random dot product graph (RDPG) model. The RDPG model provides a powerful and interpretable framework for modeling network data through low-dimensional node embeddings. It models the probability of an edge between two nodes as a dot product of their latent position vectors and is easily estimated via the well-known adjacency spectral embedding. Maximum likelihood estimation of logistic RDPGs is outlined for link prediction, community detection under partial observation, and an out-of-sample extension that makes the model suitable for time-varying networks. Empirical studies on both simulated and real-world datasets, including political blog networks and co-purchase graphs, demonstrate that the approach outperforms existing approaches.
