1

In this paper, we identify the problem of limited learnable information problem in graph fairness learning. To address this problem, we present G-Fame, a novel plug-and-play method for assisting any GNNs to learn discriminative representations. Specifically, G-Fame introduces an MoE mechanism that utilizes multiple expert neural networks to capture different aspects of knowledge in the fairness setting. In addition, we propose G-Fame++ with three innovative regularization strategies to further increase the diversity from perspectives of node representations, layers, and experts. Extensive experiments and in-depth studies demonstrate the superiority of G-Fame and G-Fame++ across a variety of accuracy and fairness metrics on multiple benchmark datasets.

In this paper, we first identify the fundamental issue in adversarial graph learning: the mixture dis- tribution between clean and attacked data samples. Motivated by this problem, we propose Graph Adversarial Mixture of Experts (GAME), a novel method to improve the model capacity, augment adversarial graphs, and enrich the graph representation diversity. For acquiring these triple im- provements, GAME contains three innovative components, including a plug-and-play GAME layer, a decoupling graph adversarial training strategy DECOG, and a graph diversity regularization strat- egy GRADIV. GAME outperforms other baselines when evaluated on several datasets given different attack methods. Additional experimental analyses prove the effectiveness of GAME in handling the complex mixture distribution, generating distinct adversarial graphs, and learning distinguishable representations.

In this paper, we reveal the necessity of malicious behavior detection on social media and propose a dataset with both user and tweet labels for this research. To integrate both structural and semantic information, we create a graph with multiple nodes including users, tweets, mentions, hashtags, topics, and multiple-name entities. Based on the graph data, we introduce a novel GraphBERT model which integrates both semantic and relational information.

In this paper, we propose a data-driven approach that trains regression models for forecasting the state-level COVID-19 daily transmission rates using the state-level mobility data and restrictive policies.