Comparing Graph Embedding Features to Social Structural Features
Jesse FaganGraph embedding methods are a new neural network method of analyzing social networks where the researcher may not know a priori which structural features of a network are important to measure. They use an approach similar to convolutional neural networks which inductively discover the features of images which best help in a prediction or classification task. The goal of this work is to encourage broader understanding of current developments in neural network graph-embedding methods (specifically node2vec and DeepWalk) and how they compare to the multitude of existing successful approaches to generating and measuring features of networks. I will use three different existing network datasets (one whole network survey, one email network, and one a set of ego networks) where structural features such as centrality, diversity, and structural holes have been known to predict outcomes such as turnover and organizational attachment. I will then show how two neural network methods, node2vec and DeepWalk, produce structural features of the nodes in each network, and I will compare how well these features perform relative to one another to predict individual outcomes. The results will help social network researchers understand how these methods produce results and how to understand their outputs.