Abstract
Social media platforms face an ongoing battle against evolving spamming tactics. Traditional detection methods often fail to account for the varying organizational scales of spammers, from large botnets to independent actors. This paper presents SGASD (Sparse Group Modeling for Adaptive Spammer Detection), a framework that utilizes a hierarchical index tree and sparse learning to adaptively identify suspicious group structures. By combining social network analysis with message content, SGASD achieves superior detection performance and robustness, particularly in identifying spammers that operate outside of dense clusters.
Introduction
Spammer detection typically relies on link-based, neighbor-based, or group-based analysis. However, modern spammers employ "link farming" to deceive these models and vary their group sizes to avoid detection. The core challenge lies in determining the optimal group granularity—identifying whether a spammer is part of a small cell, a massive network, or acting alone. Existing methods often struggle with this variance, either focusing too narrowly on individuals or too broadly on fixed clusters.
Methodology
The SGASD framework addresses these challenges by treating spammer detection as a sparse learning task. It employs Adaptive Group Structure Modeling, using the Louvain method to construct a hierarchical index tree. This tree represents users at multiple resolutions, from individual leaf nodes to small clusters and large root communities.
To isolate relevant groups, the model applies a Sparse Group Selection mechanism using a group lasso-style penalty ($\ell_{1,2}$-norm). This mathematical constraint filters out noise and focuses on the most discriminative structures. Furthermore, SGASD integrates message content analysis with network structure through an integrated objective function, optimized via Stochastic Gradient Descent (SGD) for scalability on large-scale social media data.
Results and Insights
The model was evaluated on two Twitter datasets: TwitterS (suspended accounts, 7.5% spammers) and TwitterH (honeypot accounts, 50% spammers). SGASD consistently outperformed state-of-the-art baselines, achieving an $F_1$ score of 88.01% on TwitterS and 95.31% on TwitterH.
A key advantage of SGASD is its ability to detect "sparse" spammers. While baseline methods like SSDM primarily identify spammers in dense, obvious clusters, SGASD's hierarchical approach allows it to find spammers in sparse areas of the feature space—the so-called "lone wolves" who detach from groups to avoid detection.
Conclusion
SGASD provides a robust and scalable solution for large-scale spammer detection by adaptively modeling social structures. By leveraging sparsity as a filter and integrating content with network data, it effectively addresses the limitations of fixed-granularity models. Future work will explore incorporating heterogeneous data sources like timestamps and location data to further enhance detection accuracy against increasingly subtle spamming tactics.
Summary updated