Abstract: Detecting network spammers is crucial for creating a harmonious online environment. Existing graph Transformer-based methods for network spammers detection face challenges due to indiscriminate information propagation between nodes within communities. This leads to overly homogeneous node representations and issues with excessive compression and smoothing when handling long-range dependencies, ultimately reducing the effectiveness of network spammers detection. A novel model, the relational bi-directional graph Mamba Fourier Kolmogorov-Arnold network (RMFKAN), is proposed to address these challenges in detecting network spammers on social platforms. The method of heterogeneous perception long-distance relationship feature extraction is used to solve the problem of feature loss in long-distance relationships across communities in large-scale social networks. The bi-directional selection state space model (Bi-Mamba) is introduced to address the issues of over-compression and over-smoothing when dealing with long-distance dependencies. Specifically, subgraphs are tokenized by the random walk strategy, message passing neural networks are input to independently handle different types of edges, and features are enhanced by KAN improved with Fourier coefficients. The feature matrix is input into Bi-Mamba to improve the ability of capturing long-distance dependencies and effectively reduce training complexity. On the two public online spammer detection datasets Twibot-20 and Twibot-22, compared with 10 baseline models, the experimental results show that RMFKAN is superior to existing baseline methods in multiple evaluation indicators. Compared with the best results of existing research, the F1 score of RMFKAN is increased by 2.10 and 4.06 percentage points respectively, and the accuracy is increased by 1.01 and 4.45 percentage points respectively, which verifies its superior performance in the task of network spammers detection.

Key words: network spammers detection, graph neural network, random walk, Mamba

摘要： 网络水军检测任务对构建和谐网络空间至关重要。针对现有基于图Transformer的网络水军检测方法无差别传递来自社区的节点之间的信息，从而导致节点表示过于同质，在处理长距离依赖关系时存在过度压缩和过度平滑，最终降低网络水军检测效果的问题，提出了一种基于关系双向图Mamba的傅里叶Kolmogorov-Arnold网络（RMFKAN）模型用于检测社交平台中的网络水军。通过异质感知的长距离关系特征提取方法解决了大规模社交网络跨社区长距离关系特征丢失的问题。通过引入双向选择状态空间模型（Bi-Mamba）解决了处理长距离依赖关系时的过度压缩和过度平滑问题。具体而言，通过随机游走策略令牌化子图，输入消息传递神经网络独立处理不同类型的边，利用傅里叶系数改进的KAN增强特征，将特征矩阵输入Bi-Mamba，提高对长距离依赖关系的捕捉能力，同时有效降低训练复杂度。在两个公开的网络水军检测数据集Twibot-20和Twibot-22上与10个基线模型进行对比实验，实验结果表明，RMFKAN在多个评价指标上均优于现有的基线方法，与现有研究的最佳效果相比RMFKAN的F1分数分别提高了2.10和4.06个百分点，准确率分别提高了1.01和4.45个百分点，验证了其在网络水军检测任务中的优越性能。

关键词: 网络水军检测, 图神经网络, 随机游走, Mamba