RABot: Reinforcement-Guided Graph Augmentation for Imbalanced and Noisy Social Bot Detection
Authors: Longlong Zhang, Xi Wang, Haotong Du, Yangyi Xu, Zhuo Liu, Yang Liu
Organizations: School of Artificial Intelligence, Optics and Electronics (iOPEN), Northwestern Polytechnical University · Shenzhen Research Institute, Northwestern Polytechnical University · School of Automation Science and Engineering, Xi’an Jiaotong University · School of Computer Science, Northwestern Polytechnical University
Social bot detection is pivotal for safeguarding the integrity of online information ecosystems. Although recent graph neural network (GNN) solutions achieve strong results, they remain hindered by two practical challenges: (i) severe class imbalance arising from the high cost of generating bots, and (ii) topological noise introduced by bots that skillfully mimic human behavior and forge deceptive links. We propose the Reinforcement-guided graph Augmentation social Bot detector (RABot), a multi-granularity graph-augmentation framework that addresses both issues in a unified manner. RABot employs a neighborhood-aware oversampling strategy that linearly interpolates minority-class embeddings within local subgraphs, thereby stabilizing the decision boundary under low-resource regimes. Concurrently, a reinforcement-learning-driven edge-filtering module combines similarity-based edge features with adaptive threshold optimization to excise spurious interactions during message passing, yielding a cleaner topology. Extensive experiments on three real-world benchmarks and four GNN backbones demonstrate that RABot consistently surpasses state-of-the-art baselines. In addition, since its augmentation and filtering modules are orthogonal to the underlying architecture, RABot can be seamlessly integrated into existing GNN pipelines to boost performance with minimal overhead.
Figures & tables
Figure 1: Unreliable aggregation in social networks (left). Performance variation after removing all unreliable edges on Twibot-20 (right).
Figure 2: Overall structure of RABot model, which consists of four modules: user information representation module, feature augmentation module, reinforcement edge filtering module and GNN classification module.
Methods
Dataset
Cresci-15
Twibot-20
MGTAB
Metrics
Accuracy
F1-score
Accuracy
F1-score
Accuracy
F1-score
Feature-based
Wei et al.
96.10 ± 0.75
81.68 ± 0.68
70.23 ± 0.17
53.37 ± 0.16
-
-
Varol et al.
93.25 ± 0.24
94.82 ± 0.10
78.73 ± 0.37
81.26 ± 0.42
-
-
SATAR
93.14 ± 0.51
94.87 ± 0.32
84.50 ± 0.67
86.54 ± 0.68
-
-
SGBot
77.40 ± 0.82
78.15 ± 0.13
81.63 ± 0.55
84.91 ± 0.34
-
-
Graph-based
GCN
96.49 ± 0.37
96.32 ± 0.20
77.62 ± 0.85
80.36 ± 0.49
83.61 ± 1.47
78.16 ± 1.65
Table 1: Performance comparison of varied detection methods on the Cresci-15, Twibot-20, and MGTAB datasets. Each method is executed five times, and we report the mean and standard deviation (mean ± std%). “GNN (+) ” denotes the absolute improvement of RABot (GNN) over its backbone GNN, whereas “–” indicates that the method cannot be applied to MGTAB. The best and second-best scores are highlighted with bold and underline , respectively.
Figure 3: Performance of RABot with a RGCN backbone versus random edge removal of different drop rates on the Cresci-15, Twibot-20, and MGTAB datasets.
Figure 4: Data-efficiency study. The model is trained on random subsamples of the original training data ( 10% - 100% ) from Twibot-20 and MGTAB, and evaluates on the full test sets.
Ablation
Twibot-20
MGTAB
Accuracy
F1-score
Accuracy
F1-score
RGT backbone
RABot
87.92 ± 0.12
88.40 ± 0.28
90.74 ± 0.15
88.76 ± 0.16
w/o MA
87.40 ± 0.27
88.08 ± 0.17
90.42 ± 0.31
88.41 ± 0.29
w/o FA
87.31 ± 0.16
87.96 ± 0.31
89.90 ± 0.19
86.88 ± 0.23
w/o EF
86.96 ± 0.12
87.74 ± 0.34
89.63 ± 0.18
86.96 ± 0.37
Table 2: Ablation study on the Twibot-20 and MGTAB datasets. “w/o” indicates that the specified component is removed. MA, FA, EF, and GC denote the multi-head attention, feature-augmentation, enhanced edge-filtering, and GNN classification modules, respectively.
Thresholds
Twibot-20
MGTAB
Accuracy
F1-score
Accuracy
F1-score
0.2
86.16
86.04
89.97
87.33
0.4
86.62
86.48
90.54
88.15
0.6
86.73
86.61
90.32
87.84
0.8
86.06
85.94
89.89
87.10
ours
87.36
87.12
91.16
89.03
Table 3: Performance comparison between fixed and dynamic filtering thresholds (ours) on the Twibot-20 and MGTAB datasets.
Figure 5: Sensitivity of RABot with a GAT backbone to the loss weights λs and λe on the Twibot-20 and MGTAB.
Methods
Epoch Time
Epoch
Training Time
Peak Memory
BECE
8.4s
100
14min
8268M
LMBot
4.2s
300
21min
18636M
BotDGT
27s
20
9min
5724M
SEBot
13.5s
80
18min
14482M
RABot
1.0s
300
5min
2614M
Table 4: Comparisons of computational cost on Twibot-20.
School of Cyber science at University of Science and Technology of China, Hefei 230026, China · State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China · Institute of Data Science at the National University of Singapore, 119077, Singapore +4