As Large Language Models (LLMs) demonstrate impressive capabilities, demystifying their internal mechanisms becomes increasingly vital. Neuron attribution, which attributes LLM outputs to specific neurons to reveal the semantic properties they learn, has emerged as a key interpretability approach. However, while neuron attribution has made significant progress in deciphering text-only LLMs, its application to Multimodal LLMs (MLLMs) remains less explored. To address this gap, we propose a novel Neuron Attribution method tailored for MLLMs, termed NAM. Specifically, NAM not only reveals the modality-specific semantic knowledge learned by neurons within MLLMs, but also highlights several intriguing properties of neurons, such as cross-modal invariance and semantic sensitivity. These properties collectively elucidate the inner workings mechanism of MLLMs, providing a deeper understanding of how MLLMs process and generate multi-modal content. Through theoretical analysis and empirical validation, we demonstrate the efficacy of NAM and the valuable insights it offers. Furthermore, leveraging NAM, we introduce a multi-modal knowledge editing paradigm, underscoring the practical significance of our approach for downstream applications of MLLMs.

Graph representation learning on vast datasets, like web data, has made significant strides. However, the associated computational and storage overheads raise concerns. In sight of this, Graph condensation (GCond) has been introduced to distill these large real datasets into a more concise yet information-rich synthetic graph. Despite acceleration efforts, existing GCond methods mainly grapple with efficiency, especially on expansive web data graphs. Hence, in this work, we pinpoint two major inefficiencies of current paradigms: (1) the concurrent updating of a vast parameter set, and (2) pronounced parameter redundancy. To counteract these two limitations correspondingly, we first (1) employ the Mean-Field variational approximation for convergence acceleration, and then (2) propose the objective of Gradient Information Bottleneck (GDIB) to prune redundancy. By incorporating the leading explanation techniques (e.g., GNNExplainer and GSAT) to instantiate the GDIB, our EXGC, the Efficient and eXplainable Graph Condensation method is proposed, which can markedly boost efficiency and inject explainability. Our extensive evaluations across eight datasets underscore EXGC's superiority and relevance. Code is available at https://github.com/MangoKiller/EXGC.

Graph anomaly detection (GAD) has various applications in finance, healthcare, and security. Graph Neural Networks (GNNs) are now the primary method for GAD, treating it as a task of semi-supervised node classification (normal vs. anomalous). However, most traditional GNNs aggregate and average embeddings from all neighbors, without considering their labels, which can hinder detecting actual anomalies. To address this issue, previous methods try to selectively aggregate neighbors. However, the same selection strategy is applied regardless of normal and anomalous classes, which does not fully solve this issue. This study discovers that nodes with different classes yet similar neighbor label distributions (NLD) tend to have opposing loss curves, which we term it as "loss rivalry". By introducing Contextual Stochastic Block Model (CSBM) and defining NLD distance, we explain this phenomenon theoretically and propose a Bi-level optimization Graph Neural Network (BioGNN), based on these observations. In a nutshell, the lower level of BioGNN segregates nodes based on their classes and NLD, while the upper level trains the anomaly detector using separation outcomes. Our experiments demonstrate that BioGNN outperforms state-of-the-art methods on four benchmarks and effectively mitigates "loss rivalry".

Causal effect estimation from networked observational data encounters notable challenges, primarily hidden confounders arising from network structure, or spillover effects that influence unit's outcomes based on neighboring treatment assignments. Existing graph neural network (GNN)-based methods have endeavored to address these challenges, utilizing the GNN's message-passing mechanism to capture hidden confounders or model spillover effects. However, they mainly focus on transductive causal effect learning on a single networked data, limiting their efficacy in inductive settings for real-world applications where networked data often originates from multiple environments influenced by potentially varying time or geographical regions. In light of this, we introduce the principle of invariance to the task of causal effect estimation on networked data, culminating in our Invariant Graph Learning (IGL) framework. Specifically, it first generates multiple networked data to simulate diverse environments from a given observational data. Then it further encourages the model to learn environment-invariant representations for confounders and spillover effects. Such a design enables the model to extrapolate beyond a single observed environment, thereby improving the performance of causal effect estimation in potential new environments. Extensive experiments on two real-world datasets demonstrates the superiority of our approach.

Graph anomaly detection (GAD) under semi-supervised setting poses a significant challenge due to the distinct structural distribution between anomalous and normal nodes. Specifically, anomalous nodes constitute a minority and exhibit high heterophily and low homophily compared to normal nodes, which makes the distribution of neighbors of the two types of nodes close, that is, most of them are composed of normal nodes, which causes the two types of nodes to be difficult to distinguish during the aggregation process. Furthermore, we discover that apart from various time factors and annotation preferences, graph adversarial attacks can lead to and amplify the heterophily difference across training and testing data, which is called structural distribution shift (SDS) in this paper. Current mainstream methods for GAD tend to overlook the SDS problem, resulting in poor generalization performance and limited effectiveness in detecting anomalies. This work solves the problem from a feature view. We observe that the degree of SDS varies between anomalies and normal nodes. Hence to address the issue, the key lies in resisting high heterophily for anomalies meanwhile benefiting the learning of normals from homophily. Since different labels correspond to the difference of critical anomaly features which make great contributions to the GAD, we tease out the anomaly features on which we constrain to mitigate the effect of heterophilous neighbors and make them invariant. However, the prior distribution of anomaly features is dynamic and hard to estimate, we thus devise a prototype vector to infer and update this distribution during training. For normal nodes, we constrain the remaining features to preserve the connectivity of nodes and reinforce the influence of the homophilous neighborhood. We term our proposed framework asGraphDecompositionNetwork(GDN). To demonstrate the effectiveness of the network, we explain the process of feature decomposition in the spectral domain. Extensive experiments are conducted on four benchmark datasets, including two additional datasets and two used in the preliminary work. To further validate our performance under SDS, we conduct an adversarial attack to incur different heterophily degrees for the training set and the test set. The proposed framework achieves remarkable accuracy and robustness boost in GAD, especially in an SDS environment where anomalies have largely different structural distribution across training and testing environments. Our code is open-sourced inhttps://github.com/fortunato-all/skl-GDN.

Machine learning models are increasingly applied to loan default prediction to reduce the labor cost of financial institutions and the waiting time of lenders. We find that existing loan default prediction models remain lack minimax fairness, i.e., encountering significant performance drops on underrepresented subpopulations. The main cause of this trustworthy issue is pursuing Empirical Risk Minimization over the whole population, which will overlook the underrepresented subpopulations. To tackle this issue, we split the training data into subpopulations (a.k.a. environments) and conduct Invariant Risk Minimization (IRM) to learn the optimal prediction model across environments. A technical challenge is the computation cost of directly using existing IRM methods suitable for loan default prediction, such as meta-IRM, which quadratically increases as the number of environments. To reduce the complexity in training, we propose a light meta-IRM method which reduces time complexity to be linear through environment sampling and loss replaying strategies. We apply the light meta-IRM to train a representative loan default prediction model and conduct both online and offline evaluations on a large auto loan platform. Extensive experiment results validate the advantage of the proposed light meta-IRM w.r.t. the overall accuracy, minimax fairness, and training cost.

This work studies the evaluation of explaining graph neural networks (GNNs), which is crucial to the credibility of post-hoc explainability in practical usage. Conventional evaluation metrics, and even explanation methods -- which mainly follow the paradigm of feeding the explanatory subgraph and measuring output difference -- always suffer from the notorious out-of-distribution (OOD) issue. In this work, we endeavor to confront the issue by introducing a novel evaluation metric, termed **O**OD-resistant **A**dversarial **R**obustness (OAR). Specifically, we draw inspiration from the notion of adversarial robustness and evaluate post-hoc explanation subgraphs by calculating their robustness under attack. On top of that, an elaborate OOD reweighting block is inserted into the pipeline to confine the evaluation process to the original data distribution. For applications involving large datasets, we further devise a **Sim**plified version of **OAR** (SimOAR), which achieves a significant improvement in computational efficiency at the cost of a small amount of performance. Extensive empirical studies validate the effectiveness of our OAR and SimOAR.

Graph anomaly detection (GAD) is a challenging binary classification problem due to its different structural distribution between anomalies and normal nodes --- abnormal nodes are a minority, therefore holding high heterophily and low homophily compared to normal nodes. Furthermore, due to various time factors and the annotation preferences of human experts, the heterophily and homophily can change across training and testing data, which is called structural distribution shift (SDS) in this paper. The mainstream methods are built on graph neural networks (GNNs), benefiting the classification of normals from aggregating homophilous neighbors, yet ignoring the SDS issue for anomalies and suffering from poor generalization.

Graph anomaly detection (GAD) suffers from heterophily — abnormal nodes are sparse so that they are connected to vast normal nodes. The current solutions upon Graph Neural Networks (GNNs) blindly smooth the representation of neiboring nodes, thus undermining the discriminative information of the anomalies. To alleviate the issue, recent studies identify and discard inter-class edges through estimating and comparing the node-level representation similarity. However, the representation of a single node can be misleading when the prediction error is high, thus hindering the performance of the edge indicator.