Network Embedding Methods for Large Networks in Political Science

• In Event: Measurement Using Online Data

Abstract

Social networks play an important role in many studies of political science. These networks can reflect interactions between groups of actors at the individual, national, and international levels including social media users, interethnic groups, politicians, news media, and more. Social and political interactions can be modelled as a network, where the nodes that represent persons or institutions (political actors, news sources, social media users, or social groups), that are linked together by edges that represent actions (online following behavior, sharing behavior, offline interactions, or other actions that link individuals or institutions of interest). Analysis of these networks--due to their size, sparsity and complexity--often requires generation of feature representations that can be used in machine learning models. One way to generate these feature representations is to use network embedding methods for learning low-dimensional feature representations of nodes and edges in a network. While there is some literature on comparison of advantages and shortcomings of these models, to our knowledge, there has not been any analysis on the applicability of network embedding models in different classification tasks in political science. We aim to fill this gap in the literature.

Previous research has demonstrated the power of network-features in several applications, for example Barbera et al. describe a user-centric ideal point estimation that relies on the network of friends and followers of politicians on Twitter, a network that has hundreds of millions of nodes and edges, that can accurately estimate user ideology (Barbera et al. 2015). Likewise, efforts to estimate the ideology of media domains relies on the network of Twitter users and the domains they share on Twitter, or the media domains and the subreddits they are shared in on Reddit (Eady et al. 2021, Brown et al. 2021). Additionally, clustering algorithms to predict whether or not a domain is a source of fake or misleading news relies on the large network of users and the domains they share on social media platforms (Grinberg et al. 2019).

Existing methods for dealing with these networks typically use discretized adjacency matrices, and require removing potentially valuable information prior to downstream classification tasks. Furthermore, the high dimensionality of these networks can lead to a variety of negative consequences when used for training data in machine learning models, including but not limited to reduced accuracy, overfitting, and slow training. Network embedding methods, however, provide a dense, more reliable, and more useful numerical space, thus facilitating the subsequent classification task. As a result, these models allow a researcher to classify the larger number of nodes with potentially higher accuracy and lower computational costs.

In this paper, we investigate five different feature representations for networks: Principal Component Analysis (PCA), Node2Vec, Large-scale Information Network Embedding (LINE), Scalable Incomplete Network Embedding (SINE), and GraphSAGE. We use each of these methods for the prediction of three specific variables of interest that represent a range of typical political science tasks. These are 1) the ideology of users on Reddit and Twitter, 2) ideology of domains on Reddit and Twitter, and 3) domains commonly shared on Reddit and Twitter that are likely to produce fake news. We compare the percentage of the network that we can embed, the accuracy of the labels produced for nodes, and the computational costs of each of these methods. Finally, we provide researchers with guidelines for choosing the best type of network embedding method based on network structure and the type of classification task they are performing.

Authors

• Megan Brown, New York University

• Zhanna Terechshenko, New York University

• Rachel Connolly, Center for Social Media and Politics

• Angela Lai, New York University

• Tianxin Ji, New York University

• Jonathan Nagler, New York University

• Joshua A. Tucker, New York University

• Richard Bonneau, NYU