Graph Neural Networks for Antisocial Behavior Detection on Twitter

With the rise of social media platforms, interpersonal communication has become easier and more frequent. However, antisocial behavior has also experienced an increase in various forms such as stereotypical or hateful comments toward individuals or social groups, false or distorted news, aggression, violence, etc. Although it could be beneficial for the author in terms of reaching more audiences or getting more views, likes, etc., it can be harmful to the target. Being able to detect online antisocial behavior could be a significant asset for social media platforms that enable them to perform actions to prevent it.

Graph Neural Networks (GNNs) are deep learning-based models that operate on graph structures. GNNs learn embedding representation for each node in the graph. Edge embeddings and graph embeddings can be created with the aggregation of node embeddings. GNNs perform two operations on the node embeddings obtained by the previous layer and the adjacency matrix of the graph [1, 2, 3, 4]. The first operation is graph filtering which computes node embeddings, while the second is graph pooling which generates a smaller graph with fewer nodes and its corresponding new node embeddings. There is a variety of GNN models that implement various graph filtering functions.

In the past few years, GNNs have gained interest in the Natural Language Processing (NLP) field for text classification [5, 6]. The traditional models based on recurrent neural networks (RNNs), convolutional neural networks (CNNs), and/or transformers capture contextual (local) information within a sentence. On the other hand, graph-based approaches capture global information about the vocabulary of a language [6]. Since the text data does not naturally have a graph structure, the crucial and most important part is to represent the text as a graph. Early approaches are focused on constructing text graphs composed of word nodes and documents nodes [5], while more recent approaches demonstrate that augmenting with additional information such as part of speech (POS) tags, named entities, and transformer-based word/sentence embeddings is beneficial.

In this paper, we evaluate the performance of several graph neural networks on the problem of detecting fake news and hate speech spreaders on Twitter¹¹1The code for this research is available at: https://github.com/mtoshevska/Antisocial-Behavior-on-Twitter. We define the problem as a node classification problem. We have created heterogeneous graphs using the datasets provided by a series of shared tasks on digital text forensics and stylometry (PAN) and we have trained several graph neural network models to classify user nodes. For comparison we have evaluated the proposed models on two additional tasks i.e. irony/stereotype spreaders on Twitter and sentiment classification on Yelp reviews. The rest of the paper is organized as follows. In Section 2, a brief introduction to GNN approaches to text classification problems is presented. The datasets are described in detail in Section 3. Section 4 presents the baseline models. The heterogeneous graph creation process is presented in Section 5, while Section 6 describes graph neural network models. The results are presented in Section 7. Section 8 concludes the paper.

Graph-based approaches have been evaluated for many text classification tasks. TextGCN [5] operates on a heterogeneous graph created from text data representing words and documents as nodes, and relations between them as edges. Two-layer graph convolutional network (GCN) is applied on the heterogeneous text graph to allow indirect message passing between document nodes. TextGCN significantly outperforms baseline RNN-/CNN-based models on several benchmark datasets for sentiment classification, newsgroup classification, medical abstract classification, etc. The heterogeneous graph in our study was created following the TextGCN process of graph creation.

VGCN-BERT [6] augments a BERT-based text classification model with graph embeddings to include global information about the vocabulary. A vocabulary graph has been constructed using normalized point-wise mutual information (NPMI). Vocabulary GCN (VGCN) has been applied to the vocabulary graph to create a graph embedding for the sentence. VGCN captures the part of the graph relevant to the input and then performs 2 layers of convolution, combining words from the input sentence with their related words in the vocabulary graph. To obtain the final class prediction, multiple layers of attention mechanism have been applied to the concatenated representation of the input text created with BERT and graph embeddings created with VGCN. VGCN-BERT has been evaluated on multiple text classification tasks including sentiment classification, hate speech detection, etc. In [7], a heterogeneous graph has been constructed following TextGCN [5], but a BERT/RoBERTa model has been used to obtain embeddings for the initial representation of the document nodes. The proposed model, BertGCN, has been optimized jointly with an auxiliary classifier that directly operates on BERT embeddings because it led to faster convergence and better performances. BertGCN parameters have been initialized with parameters of a pre-trained BERT model on the target dataset to speed up the training. Compared with the traditional BERT/RoBERTa models, the BertGCN yielded better performances. BertGCN has been evaluated on the same benchmark datasets as TextGCN. The performance gains obtained by BertGCN were higher for datasets containing longer sentences that enable capturing longer-term dependencies. Node representation in our study follows the BertGCN idea of document representation. We have utilized a BERT-based model to create an embedding for the initial representation of each tweet.

PAN²²2https://pan.webis.de/, last visited: 25.02.2023 is a series of scientific events and shared tasks on digital text forensics and stylometry. There is a series of author profiling shared tasks that each year are focused on a different topic. In the past three years, they were focused on antisocial behavior detection on Twitter. The participants have used a wide variety of models starting from traditional machine learning models to Transformer-based architectures [8, 9, 10]. Most of the participants have used traditional machine learning approaches with various features such as n-grams, term frequency-inverse document frequency (TF-IDF), lexicons, word embeddings, sentence embeddings, etc. A few of the participants in 2020 [8] have created deep learning models such as multi-layer perceptron (MLP), CNNs, and RNNs. In the 2021 shared task [9], one of the participants built a BERT-based model with additional linear layers; and in 2022 [10], a graph convolutional neural network was first implemented by one of the participants. The best performing model for the task of fake news spreaders detection was a Logistic Regression model trained with n-gram features, as well as some statistic-based features from the tweets such as average length or lexical diversity [11]. The best performing model for the task of hate speech spreaders detection was a CNN model that used 100-dimensional word embedding vectors [12]. The best performing model for the task of detecting irony and stereotype spreaders was a CNN model with BERT-based tweet features [13]. In our experiments, we have used datasets provided by the PAN shared tasks. Since there was only one participant utilizing GNNs for the shared tasks, we aim to investigate in detail the performance of GNNs on these datasets.

The datasets used for the experiments are provided by PAN for the Author Profiling shared tasks for the years 2020 (Profiling fake news spreaders on Twitter) and 2021 (Profiling hate speech spreaders on Twitter). The dataset for the 2022 shared task (Profiling irony and stereotype spreaders on Twitter - IROSTEREO) and a dataset for sentiment classification were also used to evaluate and compare the performances of the proposed models with more data.

Following the success of the Transformer architectures for many natural language processing tasks and to compare the performance of the graph neural network models, we have trained three Transformer-based models: DistilBERT [14], RoBERTa [15], and DistilRoBERTa [16]. DistilBERT learns an approximate version of BERT using a knowledge distillation technique [17, 18]. With only one-half of the layers of the original version of the BERT model, the number of parameters is reduced by 40%. DistilBERT is designed to be smaller and faster than BERT, while still retaining much of its accuracy. RoBERTa follows the original BERT architecture but has been trained with a different training procedure and on a larger corpus of text. It has been trained with dynamic masking where the masking pattern is generated every time a sequence is fed to the model, as opposed to static masking in the original BERT implementation where the same training mask was used. RoBERTa has been trained without the next sentence prediction objective, with bigger batches over more data and longer sequences. DistilRoBERTa is a combination of the former two models. It learns an approximate version of the RoBERTa model following the same training procedure as in DistilBERT.

Since the goal is to classify users based on the tweets they have posted, we have concatenated all tweets of a particular user into one representation. We have used PyTorch implementation of these models available in the Huggingface Transformers library⁷⁷7https://huggingface.co/docs/transformers/index, last visited: 25.02.2023. We initialized the weights with the pre-trained distilbert-base-uncased, roberta-base, and distilroberta-base weights for DistilBERT, RoBERTa, and DistilRoBERTa models, respectively. All models have been trained with AdamW optimizer, binary cross-entropy loss, and batch size 16. For the other hyperparameters, we have performed a hyperparameter search among a set of possible values. The optimal hyperparameters for each model and each dataset are summarized in Table 1.

We have created a heterogeneous graph dataset for classifying Twitter users, composed of three types of nodes: (1) user nodes, (2) tweet nodes, and (3) word nodes; and four types of edges: (1) user-tweet, (2) tweet-word, (3) word-word, and (4) tweet-tweet. The graph was created using all data in the subsets for training, validation, and testing. A simplified visualization of the graph is shown in Figure 1.

A vocabulary composed of the unique words in the dataset has been created. Special tokens representing user mentions, links, and hashtags have been added to the vocabulary. Rare words (words with less than 15 occurrences) have been removed and the remaining were used as word nodes.

Following the BertGCN [7] model, we utilize word and sentence embeddings to encode the nodes. Each word node is initialized with a word embedding of the corresponding word. We have used 200-dimensional GloVe [19] embedding vectors pre-trained on a Twitter dataset. The embeddings have been extracted using the Gensim library⁸⁸8https://radimrehurek.com/gensim/, last visited: 25.02.2023.

Each tweet is represented as a node initialized with a 768-dimensional sentence embedding obtained by a pre-trained DistilRoBERTa [16] model. The embeddings have been obtained using the Sentence-Transformers library⁹⁹9https://www.sbert.net/, last visited: 25.02.2023. User nodes have been initialized via the embedding representation of their tweets. Pre-trained DistilRoBERTa embeddings have been obtained for each of the 200 tweets per user. The embeddings have been averaged along the 0-axis thus ending with a 768-dimensional representation for each user.

Word-word and tweet-word edges have been added following the graph creation process for the TextGCN model [5]. Edges between a pair of word nodes are added if the PMI is greater than 0. PMI value has been set as a weight for word-word edges. Edges between words and tweets are added with the TF-IDF of the word in the tweet as a weight for the edge. User-tweet edges have been added between each user and their 200 tweets. Tweet-tweet edges have been added following the CLHG [20] model. Each tweet is linked with the K most similar tweets according to cosine similarity (we set the value for K to 3). The cosine similarity was computed on the corresponding 768-dimensional DistilRoBERTa sentence embeddings.

The total number of nodes and edges for each dataset is summarized in Table 2.

In this research, three GNN architectures have been investigated for antisocial behavior detection: GraphSAGE [3], Graph Attention Network (GAT) [21], and Graph Transformer [22, 23]. GraphSAGE is an inductive methodology for graph representation learning using sampling and aggregation of features from a node’s local fixed-size neighborhood. Different tasks and problems are likely to leverage different aggregation functions (e.g., mean, LSTM pooling) and/or loss functions. GAT leverages masked self-attention layers in graph neural networks. The hidden representation of the nodes is computed with a self-attention mechanism that enables the nodes to attend to neighborhood features by specifying different weights for each neighbor node. Graph Transformer [22] is a generalization of the Transformer architectures for graph structures. The attention mechanism is represented as a function of the neighborhood connectivity for each node in the graph and the positional encoding is represented by the Laplacian eigenvectors. The normalization layer is replaced by a batch normalization layer. The architecture could be extended to edge feature representation. Unified Message Passing (UniMP) [23] jointly performs feature and label propagation by embedding the partially observed labels into the same space as node features. It is trained with a masked label prediction strategy inspired by BERT. We have used the modified Graph Transformer operator from the UniMP.

Our architecture is composed of a two-layer heterogeneous graph neural network followed by a ReLU activation that maps the nodes into a low-dimensional latent space. For the purpose of classifying nodes, a fully-connected layer has been added on top of the GNN model, which infers the class for the user nodes. The architecture is the same for all three models and is displayed in Figure 2.

We have used PyTorch implementation of these models available in the PyTorch Geometric library¹⁰¹⁰10https://pytorch-geometric.readthedocs.io/en/latest/, last visited: 25.02.2023. The models have been created with the implementation for homogeneous graphs, and then are transformed into models suitable for heterogeneous graphs. All models have been trained with AdamW optimizer and binary cross-entropy loss. For the other hyperparameters, we have performed a hyperparameter search among a set of possible values. The optimal hyperparameters for each model and each dataset are summarized in Table 3.

We have performed several experiments with baseline Transformer models and GNN models. For each dataset, we have trained six models with the corresponding optimal hyperparameters shown in Table 1 and Table 3. Each of the models has been trained on Quadro RTX 8000 48GB GPU.

This paper explored the performances of graph neural networks for the task of antisocial behavior detection on Twitter. Three GNN architectures (GraphSAGE, GAT, and Graph Transformer) were evaluated against four datasets composed of Twitter users and tweets that they have posted that were provided by PAN shared tasks, and one dataset composed of Yelp users and reviews that they have written that was extracted from the Yelp Open Dataset. A heterogeneous graph dataset has been created with user, tweet/review, and word nodes, as well as five types of edges between them.

An ablation study was performed to investigate which components of the heterogeneous graph have contributed the most. The results showed that the best performances are achieved when all the graph components are included, while the worst performances were obtained when the edges between tweet pairs were excluded from the graph.

Transformer-based models were also trained on the same datasets as baseline models for comparison. When compared against the baseline models, the GNN models showed inferior performance for most of the experiments. For the experiments, pre-trained Transformer-based models (DistilBERT, RoBERTa, and DistilRoBERTa) have been used. The models are pre-trained on large datasets which gives them a significant advantage. This hypothesis leads to a possible future direction which is to first pre-train GNNs on a larger dataset, and then train on the specific datasets that were used in this research.

For two of the datasets employed in this study, GNN models showed comparable performances with second best and third best Transformer-based models. These findings indicate the capability of GNN models to learn from the types of data derived from social networks that were utilized in this research. We anticipate that GNN models could be successfully applied to other text classification or even wider natural language processing or generation tasks.