A0869
Title: Classically boosted network embeddings
Authors: Joel Nishimura - Arizona State University (United States) [presenting]
Yunpeng Zhao - Colorado State University (United States)
Abstract: Network embeddings are a popular and effective preprocessing step when performing machine learning with network data. We demonstrate that standard boosting techniques, AdaBoost and Real AdaBoost can be applied to network embedding techniques to increase performance, particularly in terms of link prediction on test data in a cross-validation context. These approaches produce results competitive with other state-of-the-art embedding approaches when applied to a number of empirical networks. Additionally, we show on simulated data that Real AdaBoost can de-aggregate some networks, wherein networks created by two independent latent features can have those separate latent features inferred by different boosted rounds. Further analysis of the performance of these boosted methods shows that they retain the characteristic robustness to over-fitting as boosting methods in classical settings.