Who dislikes you in social networks? Predicting negative ties using Graph Neural Networks.

Abstract

Nowadays more and more social network data can be represented as graphs. The availability of such structures, especially in a form of a signed and directed networks bring new challenges that can be analyzed. One of the most common challenges in this field is the problem of sign prediction of the link. The main difficulty is the fact that negative links carry different meaning than positive ones. This research focuses on comparing two methods of predicting the sign of the links. The first method utilizes feature engineering approach, where node and graph specific characteristics are extracted and fed as an input to the gradient boosting model. The second method utilizes Signed Graph Convolutional Network, which focuses on extracting node representations in a low dimensional space, which are used to predict a sign of the link. In the end both methods were compared on the left out test set of randomly chosen edges using accuracy, AUC score, precision and recall. The whole experiment was carried out on the publicly available dataset, which is commonly used as the benchmark for signed network algorithms. The final scores obtained by the models were of high quality. However, the performance differ significantly among tested classes, with positive edges being more easier to predict for the models than the negative ones.