Social bots constitute a substantial fraction of active accounts on digital platforms, fundamentally threatening information authenticity and democratic discourse. Contemporary detection methods confront critical limitations: information imbalance across heterogeneous relations, computational challenges in processing massive neighborhoods, and inadequate multi-scale representation learning. We propose HIER (Heterogeneous Information Bottleneck and Expert Routing), a pioneering framework that integrates variational information theory with mixture-of-experts paradigms for social network analysis. HIER introduces relation-aware variational information bottleneck for optimal compression across relationship types, dynamic sparse expert routing that extends mixture-of-experts to edge-level graph processing, and dual-scale mutual information maximization enhancing representation discriminability through neighborhood consistency and graph-level contrastive learning. Experimental validation demonstrates HIER's superior performance across real-world datasets, establishing new benchmarks for heterogeneous social bot detection.

Social spammers pose a serious threat to online social networks by manipulating topics, promoting malicious campaigns, and spreading low-quality or deceptive content at scale. Existing detection methods, including text-based classifiers, graph neural networks, and hybrid models, still suffer from three key limitations: they usually operate at a single structural scale, they treat content and network structure as loosely coupled signals, and they often assign equal importance to all edges within the same relation type. To address these issues, we propose MHGCA, a M ulti-scale H eterogeneous G raph learning framework with C ontent–Structure A lignment for robust spammer detection. MHGCA jointly encodes motif-, community-, and global-level structures, detects structural anomaly patterns, and employs a relation-aware transformer to adaptively align content similarity with relation-specific message passing. Experiments on two real-world Twitter datasets show that MHGCA consistently outperforms strong baselines under severe class imbalance and limited supervision, and remains stable across different training ratios and hyperparameter settings. These results suggest that explicitly aligning content with multi-scale heterogeneous structures is crucial for reliable spammer detection in social networks.

Multimedia rumor content has been widely disseminated with the rise of generative technologies. Existing rumor detection approaches typically focus independently on multi-modal data (such as text and images) or social structure analysis, and only a few researchers have attempted to integrate all three modalities for comprehensive rumor detection. Due to the complexity of the relationships between these heterogeneous data, combining them effectively remains a challenge. In this work, we present a novel Graph-Enhanced Multimodal Contrastive Learning (GMCL) to integrate textual, visual, and social graph features more efficiently for rumor detection. We utilize semantic correlation to assist cross-modal contrastive learning to capture fine-grained alignment between text and image and enhance node representations through graph contrastive learning without relying on negative samples. By aligning and integrating these different representations, our method can detect rumors more accurately. Extensive experimental results show that our model outperforms current state-of-the-art methods in multimodal rumor detection.

The rapid growth of social networks has significantly increased the prevalence of spammer activities, which poses substantial challenges to user trust, experience, and public safety. Existing spam detection methods primarily rely on easily circumvented handcrafted features and fail to capture the complexities of user behavior and social relationships fully. These methods often struggle to differentiate users in highly heterogeneous networks and do not adequately address the issue of class imbalance. To overcome these limitations, we propose a novel framework, Multi-Relation Aware Heterogeneous Graph Transformer (MRHGT), which effectively integrates heterogeneous graph representation learning to capture complex social structures. Our approach introduces a novel aggregation mechanism that combines relation-aware multi-head attention, relation-specific graph convolution networks, and cross-attention feature fusion. Additionally, we incorporate a heterogeneous graph contrastive learning strategy to handle class imbalance. Extensive experiments on real-world datasets demonstrate that our method outperforms state-of-the-art approaches in accuracy and robustness.

Stance detection aims to automatically identify social users' attitudes on specific targets by analyzing their textual content and various relationships within social networks. Existing stance detection models predominantly focus on textual content. Although some stance detection methods based on graph neural networks have achieved performance improvements by integrating textual content and social relationships, they overlook the implicit polarity signs of links between nodes. Meanwhile, many real-world social networks do not provide explicit sign information for links, which limits the applicability of existing signed network representation learning methods to these scenarios. This paper proposes a novel social network stance detection model named 'Implicit Sign-enhanced Stance Detection Model with Semantic Graph Attention Network' (ISSDM-SemGAN). Firstly, we propose a semantic graph attention network (SemGAN) that leverages Direction-distance Fusion Scoring Function (DFSF) to enhance the differences between nodes, which lays the foundation for capturing implicit link signs. Additionally, we design implicit sign-enhanced stance detection network to introduce the signed link prediction task and social balance theory to facilitate understanding and learning the implicit link polarity between nodes, thereby enhancing the performance of stance detection tasks. Extensive experiments were conducted on real social network dataset. The experimental results demonstrate that ISSDM-SemGAN achieves state-of-the-art performance.

N-ary relational facts represent semantic correlations among more than two entities. While recent studies have developed link prediction (LP) methods to infer missing relations for knowledge graphs (KGs) containing n-ary relational facts, they are generally limited to transductive settings. Fully inductive settings, where predictions are made on previously unseen entities, remain a significant challenge. As existing methods are mainly entity embedding-based, they struggle to capture entity-independent logical rules. To fill in this gap, we propose an n-ary subgraph reasoning framework for fully inductive link prediction (ILP) on n-ary relational facts. This framework reasons over local subgraphs and has a strong inductive inference ability to capture n-ary patterns. Specifically, we introduce a novel graph structure, the n-ary semantic hypergraph, to facilitate subgraph extraction. Moreover, we develop a subgraph aggregating network, NS-HART, to effectively mine complex semantic correlations within subgraphs. Theoretically, we provide a thorough analysis from the score function optimization perspective to shed light on NS-HART's effectiveness for n-ary ILP tasks. Empirically, we conduct extensive experiments on a series of inductive benchmarks, including transfer reasoning (with and without entity features) and pairwise subgraph reasoning. The results highlight the superiority of the n-ary subgraph reasoning framework and the exceptional inductive ability of NS-HART.

Link prediction is a core task in graph machine learning with wide applications. However, little attention has been paid to link prediction between two group entities. This limits the application of the current approaches to many real-life problems, such as predicting collaborations between academic groups or recommending bundles of items to group users. Moreover, groups are often ephemeral or emergent, forcing the predicting model to deal with challenging inductive scenes. To fill this gap, we develop a framework composed of a GNN-based encoder and neural-based aggregating networks, namely the Mutual Multi-view Attention Networks (MMAN). First, we adopt GNN-based encoders to model multiple interactions among members and groups through propagating. Then, we develop MMAN to aggregate members' node representations into multi-view group representations and compute the final results by pooling pairwise scores between views. Specifically, several view-guided attention modules are adopted when learning multi-view group representations, thus capturing diversified member weights and multifaceted group characteristics. In this way, MMAN can further mimic the mutual and multiple interactions between groups. We conduct experiments on three datasets, including two academic group link prediction datasets and one bundle-to-group recommendation dataset. The results demonstrate that the proposed approach can achieve superior performance on both tasks compared with plain GNN-based methods and other aggregating methods.