Heterogeneity-Aware Twitter Bot Detection with Relational Graph Transformers

Authors: Shangbin Feng, Zhaoxuan Tan, Rui Li, Minnan Luo3977-3985

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our proposal outperforms state-of-the-art methods on a comprehensive Twitter bot detection benchmark. Additional studies also bear out the effectiveness of our proposed relational graph transformers, semantic attention networks and the graph-based approach in general.
Researcher Affiliation Academia Shangbin Feng1, Zhaoxuan Tan1, Rui Li2, Minnan Luo1 1School of Electronic and Information Engineering, Xi an Jiaotong University, Xi an, China 2School of Continuing Education, Xi an Jiaotong University, Xi an, China
Pseudocode Yes Algorithm 1: Model Learning Algorithm
Open Source Code Yes Our implementation is publicly available on Git Hub. 1. https://github.com/BunsenFeng/BotHeterogeneity
Open Datasets Yes Our bot detection model is graph-based and heterogeneity-aware, which requires data sets that provide certain type of graph structure. Twi Bot-20 (Feng et al. 2021c) is a comprehensive Twitter bot detection benchmark and the only publicly available bot detection dataset to provide user follow relationships to support graph-based methods. In this paper, we make use of Twi Bot-20, which includes 229,573 Twitter users, 33,488,192 tweets, 8,723,736 user property items and 455,958 follow relationships.
Dataset Splits Yes We follow the same splits provided in the benchmark so that results are directly comparable with previous works.
Hardware Specification Yes Our implementation is trained on a Titan X GPU with 12GB memory.
Software Dependencies No We use pytorch (Paszke et al. 2019), pytorch lightning (Falcon 2019), torch geometric (Fey and Lenssen 2019) and the transformers library (Wolf et al. 2020) for an efficient implementation of our proposed Twitter bot detection framework.
Experiment Setup Yes We present our hyperparameter settings in Table 2 to facilitate reproduction. Hyperparameter Value optimizer Adam W learning rate 10 3 L2 regularization λ 3 10 5 batch size 256 layer count L 2 dropout 0.5 size of hidden state 128 maximum epochs 40 transformer attention heads C 8 semantic attention heads D 8 relational edge set R {follower, following}