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.
Stance detection aims to identify individuals' attitudes toward a specific target. The signs of links are beneficial in various graph-related tasks. However, existing stance detection methods neglect this information. This is because the signs of links are not explicitly provided in many real-world social networks, which makes current signed network representation learning methods inapplicable to this scenario. In this work, we propose a Sign-Aware Graph Learning framework (SAGL) for stance detection by innovating a Sign-Inferred Graph Auto-Encoder (SIGAE) and a Balance-Infused Signed Graph Attention Network (BSGAN). SIGAE infers and generates signed links from unsigned networks by exploiting stance labels and underlying structure information. BSGAN introduces balance theory into signed graph attention networks and refines node representations by capturing and dynamically adjusting the influence of signed links. Extensive experiments on real-world datasets demonstrate that SAGL significantly outperforms competitive baselines.
Tabular anomaly detection is central to structured decision systems in finance, healthcare, and industrial monitoring, where undetected anomalies can indicate fraud, disease, or system failure.Most deep detectors learn correlations in the training data, which can make their anomaly scores unstable under distribution shift. They can also be weak for hidden mechanism violations that remain statistically plausible: local feature relationships look normal, but the coordinated cross-variable changes required by the underlying mechanism are inconsistent.Causal representation learning offers a principled route to robustness, but existing methods usually require ground-truth graphs, interventions, or strong structural priors that are rarely available for high-dimensional tabular data. Our key insight is that anomaly detection can exploit mechanism-level structure without recovering a complete causal graph, because anomalies often disrupt groups of variables governed by the same generative mechanism.We propose PRISM-VAE, a variational framework that learns latent mechanism blocks by shaping local sensitivity in the representation space, without predefined causal graphs. This shifts from variable-level structure learning to mechanism aggregation, reducing overhead to $O(m)$ and making mechanism-level modeling practical for high-dimensional tabular applications. Across twelve real-world datasets, PRISM-VAE improves AUROC by up to 11.3 percentage points over strong baselines and shows stronger robustness under evaluated distribution shifts. Using a controlled synthetic financial fraud benchmark whose data-generating process specifies the ground-truth causal mechanisms, PRISM-VAE achieves mechanism aggregation without structural priors and remains competitive with methods given causal supervision. We also introduce a mechanism-level metric to test alignment between learned latent blocks and ground-truth mechanisms.