NeighborSense: A Novel GNN Framework for Social Bot Detection

Social media platforms are increasingly plagued by malicious actors utilizing automated accounts, often referred to as social bots, to spread misinformation, manipulate public opinion, and disrupt online communities. Detecting these bots is a critical challenge for maintaining the integrity of online social networks (OSNs). Recent advancements in graph neural networks (GNNs) have shown promise in addressing this challenge by leveraging both the structural and attribute data inherent in social networks.

The Rise of GNNs in Bot Detection

GNN-based techniques have gained prominence in social bot detection due to their ability to learn from the complex relationships within social graphs. Models like BotRGCN⁵ and frameworks leveraging heterogeneous information networks (HINs)⁶ effectively combine structural and attribute data to identify bot accounts. These approaches iteratively gather information from a node’s neighborhood, improving detection accuracy.

Limitations of Existing GNN Methods

Despite their success, current GNN-based bot detection methods face limitations. A key issue is the lack of effective control over the strength of local aggregation. Existing methods often rely on point-to-point attention mechanisms based on node features, failing to account for the distribution of features within a node’s neighborhood. For example, during an online event, the features of a legitimate user can be overshadowed by a large number of bot accounts, reducing detection accuracy. Current GNN models struggle to learn and control aggregation strength effectively across edges with different directions and types, such as “followed by” versus “following.”

Introducing NeighborSense: A Modern Approach

To address these limitations, researchers have developed NeighborSense, a GNN-based social bot detection method that leverages both user attributes and structural information from multiple social relations. The framework incorporates two key metrics to dynamically update a shortcut module, learning aggregation strategies that model more refined social interaction patterns.

Key Design Goals

The development of NeighborSense was guided by three primary design goals:

• Relational Heterogeneity Awareness: The model adapts to multiple social relations and learns efficient aggregation strategies for different edge types.
• Local Entropy Awareness: The aggregation strategy adjusts based on the feature distribution within a node’s neighborhood, distinguishing between homogeneous and heterogeneous neighbors.
• Local Feature Heterogeneity Awareness: The aggregation strategy adjusts based on the degree of similarity between a central user and their neighbors.

How NeighborSense Works

NeighborSense models the social network as a graph with users as nodes and interactions as edges of multiple relation types. A Relational Graph Convolutional Network (R-GCN)⁷ serves as the backbone, providing a foundation for node classification. The framework then employs two metrics to quantify the feature distribution and homogeneity of nodes and their neighbors. An adaptive shortcut module uses these metrics to allow each node to independently aggregate neighborhood information based on its neighbors’ prior predictions.

Key Contributions

• Investigation of local interaction patterns between bot and human accounts, introducing novel metrics to characterize node-level neighborhood statistics.
• A GNN-based detection framework integrating node attributes and multiple types of social relations, with a dynamically updated shortcut module for adaptive gated aggregation.
• Demonstrated superior performance compared to state-of-the-art GNN-based bot detection methods through extensive evaluations.

Experimental Results

Evaluations on the TwiBot-20 and TwiBot-22 datasets demonstrate NeighborSense’s effectiveness. The model achieves competitive performance even with a constrained model size, highlighting its efficiency. The adaptive gating mechanism allows for more refined aggregation, improving detection accuracy compared to methods with fixed aggregation strengths. Performance remains satisfactory even when using limited labeled training data.

Future Directions

Further research could explore the application of NeighborSense to other social media platforms and investigate the use of alternative neighborhood indicators to refine the gating mechanism. Exploring the robustness of the model against adversarial attacks and developing methods for real-time bot detection are likewise promising avenues for future work.

References

1. [5] BotRGCN
2. [6] Feng et al.
3. [7] R-GCN