Identifying Author Fingerprints in Texts via Graph Neural Networks

Abstract

The world is generating more and more network data in many different areas (e.g., sensor networks, social networks and even text). A unique characteristic of these data is the coupling between data values and underlying irregular structure on which these values are defined. Thus, researchers developed Graph Neural Networks (GNNs) to use deep learning approaches on these irregular network data. GNNs developers tried to replicate the recent success of Convolutional Neural Networks (CNNs) and developed its graph counterpart Graph Convolutional Neural Network (GCNN) and more different variations of GNNs (e.g. EdgeNet). However, all these architectures are relatively young, and the impact of different parameters to classification result is not well researched compared to regular neural network architectures. To address this issue, we propose to use authorship attribution problem to research the impact of different architectures and their variations to classification accuracy and how GNNs can be used to improve on authorship attribution task compared to the baseline architectures. Explicitly, we define the dataset which is going to be used throughout the experiments and the method to convert text excerpts of authors into the network that can be classified with GNNs (called WAN). WAN is as a network that captures unique author fingerprint. We also define the set of GNN architectures (and different combinations and variations of them), baseline architecture (SVM) and experiments that are used with those architectures. This experiment setting allows us to compare different GNN architectures among themselves and the baseline architecture. Also, we define a method to reduce the dimensions of author fingerprints (WANs) and use these sparse author fingerprints for the same experiments with the same architectures. Numerical results show the improvement over the baseline architectures in nearly all defined experiments. Also, we found that more complex GNN architectures (e.g. EdgeNets) are superior to shallower architectures with more laborious experiments (e.g. classification by gender). More complex architectures also require hyperparameter re-tuning in order to achieve optimal results. Furthermore, experiments with sparse author fingerprints showed that we could achieve comparable results to standard fingerprints with faster training times and significantly reduced dimensions. GNN architectures used with sparse author fingerprints were usually superior to baseline architectures.