Towards structured NLP interpretation via graph explainers

Graph neural networks (GNNs) have made rapid developments in the recent years. Due to their great ability in modeling graph-structured data, GNNs are vastly used in various applications, including high-stakes scenarios such as financial analysis, traffic predictions, and drug discovery. Despite their great potential in benefiting humans in the real world, recent study shows that GNNs can leak private information, are vulnerable to adversarial attacks, can inherit and magnify societal bias from training data and lack interpretability, which have risk of causing unintentional harm to the users and society. For example, existing works demonstrate that attackers can fool the GNNs to give the outcome they desire with unnoticeable perturbation on training graph. GNNs trained on social networks may embed the discrimination in their decision process, strengthening the undesirable societal bias. Consequently, trust-worthy GNNs in various aspects are emerging to prevent the harm from GNN models and increase the users’ trust in GNNs. In this paper, we give a comprehensive survey of GNNs in the computational aspects of privacy, robustness, fairness, and explainability. For each aspect, we give the taxonomy of the related methods and formulate the general frameworks for the multiple categories of trustworthy GNNs. We also discuss the future research directions of each aspect and connections between these aspects to help achieve trustworthiness.

Sex classification is a major benchmark of previous work in learning on the structural connectome, a naturally occurring brain graph that has proven useful for studying cognitive function and impairment. While graph neural networks (GNNs), specifically graph convolutional networks (GCNs), have gained popularity lately for their effectiveness in learning on graph data, achieving strong performance in adult sex classification tasks, their application to pediatric populations remains unexplored. We seek to characterize the capacity for GNN models to learn connectomic patterns on pediatric data through an exploration of training techniques and architectural design choices. Two datasets comprising an adult BRIGHT dataset (N = 147 Hodgkin's lymphoma survivors and N = 162 age similar controls) and a pediatric Human Connectome Project in Development (HCP-D) dataset (N = 135 healthy subjects) were utilized. Two GNN models (GCN simple and GCN residual), a deep neural network (multi-layer perceptron), and two standard machine learning models (random forest and support vector machine) were trained. Architecture exploration experiments were conducted to evaluate the impact of network depth, pooling techniques, and skip connections on the ability of GNN models to capture connectomic patterns. Models were assessed across a range of metrics including accuracy, AUC score, and adversarial robustness. GNNs outperformed other models across both populations. Notably, adult GNN models achieved 85.1% accuracy in sex classification on unseen adult participants, consistent with prior studies. The extension of the adult models to the pediatric dataset and training on the smaller pediatric dataset were sub-optimal in their performance. Using adult data to augment pediatric models, the best GNN achieved comparable accuracy across unseen pediatric (83.0%) and adult (81.3%) participants. Adversarial sensitivity experiments showed that the simple GCN remained the most robust to perturbations, followed by the multi-layer perceptron and the residual GCN. These findings underscore the potential of GNNs in advancing our understanding of sex-specific neurological development and disorders and highlight the importance of data augmentation in overcoming challenges associated with small pediatric datasets. Further, they highlight relevant tradeoffs in the design landscape of connectomic GNNs. For example, while the simpler GNN model tested exhibits marginally worse accuracy and AUC scores in comparison to the more complex residual GNN, it demonstrates a higher degree of adversarial robustness.

Molecular machine learning (ML) has proven important for tackling various molecular problems, such as predicting molecular properties based on molecular descriptors or fingerprints. Since relatively recently, graph neural network (GNN) algorithms have been implemented for molecular ML, showing comparable or superior performance to descriptor or fingerprint-based approaches. Although various tools and packages exist to apply GNNs in molecular ML, a new GNN package, named MolGraph, was developed in this work with the motivation to create GNN model pipelines highly compatible with the TensorFlow and Keras application programming interface (API). MolGraph also implements amodule to accommodate the generation of small molecular graphs, which can be passed to a GNN algorithm to solve a molecular ML problem. To validate the GNNs, benchmarking was conducted using the datasets from MoleculeNet, as well as three chromatographic retention time datasets. The benchmarking results demonstrate that the GNNs performed in line with expectations. Additionally, the GNNs proved useful for molecular identification and improved interpretability of chromatographic retention time data. MolGraph is available at https://github.com/akensert/molgraph . Installation, tutorials and implementation details can be found at https://molgraph.readthedocs.io/en/latest/ .

The n-octanol/buffer solution distribution coefficient at pH = 7.4 (log D7.4) is an indicator of lipophilicity, and it influences a wide variety of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties and druggability of compounds. In log D7.4 prediction, graph neural networks (GNNs) can uncover subtle structure-property relationships (SPRs) by automatically extracting features from molecular graphs that facilitate the learning of SPRs, but their performances are often limited by the small size of available datasets. Herein, we present a transfer learning strategy called pretraining on computational data and then fine-tuning on experimental data (PCFE) to fully exploit the predictive potential of GNNs. PCFE works by pretraining a GNN model on 1.71 million computational log D data (low-fidelity data) and then fine-tuning it on 19,155 experimental log D7.4 data (high-fidelity data). The experiments for three GNN architectures (graph convolutional network (GCN), graph attention network (GAT), and Attentive FP) demonstrated the effectiveness of PCFE in improving GNNs for log D7.4 predictions. Moreover, the optimal PCFE-trained GNN model (cx-Attentive FP, Rtest2 = 0.909) outperformed four excellent descriptor-based models (random forest (RF), gradient boosting (GB), support vector machine (SVM), and extreme gradient boosting (XGBoost)). The robustness of the cx-Attentive FP model was also confirmed by evaluating the models with different training data sizes and dataset splitting strategies. Therefore, we developed a webserver and defined the applicability domain for this model. The webserver (http://tools.scbdd.com/chemlogd/) provides free log D7.4 prediction services. In addition, the important descriptors for log D7.4 were detected by the Shapley additive explanations (SHAP) method, and the most relevant substructures of log D7.4 were identified by the attention mechanism. Finally, the matched molecular pair analysis (MMPA) was performed to summarize the contributions of common chemical substituents to log D7.4, including a variety of hydrocarbon groups, halogen groups, heteroatoms, and polar groups. In conclusion, we believe that the cx-Attentive FP model can serve as a reliable tool to predict log D7.4 and hope that pretraining on low-fidelity data can help GNNs make accurate predictions of other endpoints in drug discovery.

Past events have revealed that widespread blackouts are mostly a result of cascading failures in the power grid. Understanding the underlining mechanisms of cascading failures can help in developing strategies to minimize the risk of such events. Moreover, a real-time detection of precursors to cascading failures will help operators take measures to prevent their propagation. Currently, the well-established probabilistic and physics-based models of cascading failures offer low computational efficiency, hindering them to be used only as offline tools. In this work, we develop a data-driven methodology for online estimation of the risk of cascading failures. We utilize a physics-based cascading failure model to generate a cascading failure dataset considering different operating conditions and failure scenarios, thus obtaining a sample space covering a large set of power grid states that are labeled as safe or unsafe. We use the synthetic data to train deep learning architectures, namely Feed-forward Neural Networks (FNN) and Graph Neural Networks (GNN). With the development of GNNs, improved performance is achieved with graph-structured data, and GNNs can generalize to graphs of diverse sizes. A comparison between FNN and GNN is made and the GNNs inductive capability is demonstrated via test grids. Furthermore, we apply transfer learning to improve the performance of a pre-trained GNN model on power grids not seen in the training process. The GNN model shows accuracy and balanced accuracy above 96% on selected test datasets not used in the training. Conversely, the FNN shows accuracy above 85% and balanced accuracy above 81% on test datasets unseen during training. Overall, the GNN model is successful in determining, if one or several simultaneous outages result in a critical grid state, under specific grid operating conditions.

Recent advances in Graph Neural Networks (GNNs) have changed the landscape of modern graph analytics. The complexity of GNN training and the scalability challenges have also sparked interest from the systems community, with efforts to build systems that provide higher efficiency and schemes to reduce costs. However, we observe that many such systems basically "reinvent the wheel" of much work done in the database world on scalable graph analytics engines. Further, they often tightly couple the scalability treatments of graph data processing with that of GNN training, resulting in entangled complex problems and systems that often do not scale well on one of those axes. In this paper, we ask a fundamental question: How far can we push existing systems for scalable graph analytics and deep learning (DL) instead of building custom GNN systems? Are compromises inevitable on scalability and/or runtimes? We propose Lotan, the first scalable and optimized data system for full-batch GNN training with decoupled scaling that bridges the hitherto siloed worlds of graph analytics systems and DL systems. Lotan offers a series of technical innovations, including re-imagining GNN training as query plan-like dataflows, execution plan rewriting, optimized data movement between systems, a GNN-centric graph partitioning scheme, and the first known GNN model batching scheme. We prototyped Lotan on top of GraphX and PyTorch. An empirical evaluation using several real-world benchmark GNN workloads reveals a promising nuanced picture: Lotan significantly surpasses the scalability of state-of-the-art custom GNN systems, while often matching or being only slightly behind on time-to-accuracy metrics in some cases. We also show the impact of our system optimizations. Overall, our work shows that the GNN world can indeed benefit from building on top of scalable graph analytics engines. Lotan's new level of scalability can also empower new ML-oriented research on ever-larger graphs and GNNs.

Graph Neural Networks (GNNs) have achieved remarkable success in recent years due to their powerful ability to model non-Euclidean data structures and complex relationships. However, as graph sizes and model complexities continue to grow, the efficient deployment of GNNs across various hardware platforms presents significant challenges. The two main computation modes in GNN inference exhibit distinct characteristics in sparsity and computational density, which results in imbalanced workload distribution and inefficient utilization of compute and memory bandwidth. Furthermore, subgraph partitioning strategies, commonly adopted to address on-chip resource constraints, frequently introduce write-back overhead and port conflicts, thereby limiting overall system throughput. To address these challenges, we propose TernaryGNNs, a high-throughput, area-efficient ternary weight GNN inference framework targeting CPU-FPGA heterogeneous platforms. First, we introduce a precision-preserving ternary quantization method that maintains model accuracy with 1.016% degradation, while achieving an average weight sparsity of 81.1% and a parameter reduction rate of 94.71%. Next, by exploiting sparsity, we reformulate GNN inference as Sparse Matrix–Dense Matrix Multiplication (SpMM) or Sparse General Matrix–Matrix Multiplication (SpGEMM) computation and propose a unified sparse-optimized processor architecture. Finally, we present a comprehensive software–hardware co-design framework that ensures adaptability to the evolving landscape of diverse GNN model architectures. Our framework supports nine mainstream GNN models. Compared to the State-of-the-Art (SOTA) general-purpose GNN processor GraphOPU, TernaryGNNs achieves an average \(2.79\times\) hardware performance improvement, \(1.70\times\) end-to-end performance improvement, and \(2.83\times\) area-efficiency improvement. Compared to the leading overlay accelerator FP-GNN, it delivers \(6.73\times\) hardware performance improvement on average.

In node classification, traditional graph neural networks (GNNs) typically assume implicit homophily, indicating that intraclass nodes are likely connected. However, real-world graphs frequently exhibit heterophily, in which interclass nodes are also commonly connected. To address this challenge, recent methods have adopted approaches such as expanding local neighborhoods and employing adaptive message aggregation to enhance the GNN performance on heterophily graphs. Nevertheless, these methods are restricted by the homophily assumption and fail to effectively capture long-range dependencies (e.g., widely separated intraclass nodes) and insufficiently leverage the graph topology. This study investigates the performance differences of GNN when it is applied to both homophily and heterophily graphs and finds that the distinguishability of neighborhood label distributions (NLDs) exhibits a significant correlation with the accuracy of node classification. To assess the impact of NLD on node classification, this study proposes a novel homophily metric based on node distinguishability. Subsequently, this study introduces a new GNN model named NLD-based GNN (NLDGNN) for node classification. First, NLDGNN initializes node representations by integrating node features with node NLDs. To address long-range dependencies in heterophily graphs, NLDGNN utilizes the global label relationship matrix with low-rank characteristics for global message passing. By combining the attention scores derived from the initial node representations, NLDGNN constructs the global label relationship matrix for enhanced message passing, thereby improving the expressiveness of node representations. Experimental results indicate that NLDGNN outperforms existing GNN models on both real-world homophily and heterophily graphs. The code of this study is available at https://github.com/wanli6/NLDGNN.

Graph neural networks (GNNs) have achieved remarkable performance in many graph analytics tasks such as node classification, link prediction and graph clustering. Existing GNN systems (e.g., PyG and DGL) adopt a tensor-centric programming model and train GNNs with manually written operators. Such design results in poor usability due to the large semantic gap between the API and the GNN models, and suffers from inferior efficiency because of high memory consumption and massive data movement. We demonstrateSeastar, a novel GNN training framework that adopts avertex-centric programming paradigm and supportsautomatic kernel generation, to simplify model development and improve training efficiency. We will (i) show how to express GNN models succinctly using a visual drag-and-drop'' interface or Seastar's vertex-centric python API; (ii) demonstrate the performance advantage of Seastar over existing GNN systems in convergence speed, training throughput and memory consumption; and (iii) illustrate how Seastar's optimizations (e.g., operator fusion and constant folding) improve training efficiency by profiling the run-time performance.

Learning node representations against perturbations

Most distribution system operators may not accurately record or completely maintain the phase connections for the numerous LV customers. Different consumer phase identification (CPI) approaches based on voltages, powers or other measurements are proposed in the literature. Due to the technical challenges in collecting voltage measurements, power measurement based approaches are preferable. Hence, this paper proposes a novel power based CPI methodology applying Graph Neural Networks (GNNs). The CPI methodology generates synthetic transformer power profiles assuming random combinations of phases for the measured load profiles, which are used altogether to train the GNN model. The GNN model is then tested using measured transformer and load power profiles. The performance of the methodology is evaluated in a test low voltage grid of 55 loads under various conditions of Photovoltaic penetration, Photovoltaics under maintenance, measurement errors, unmetered consumption, uncertain grid asset parameters and inaccurate phase connections. Further tests on a real low voltage grid with 111 loads prove the scalability of the methodology. The attained results show that the GNN model can achieve accuracy above 90% in most cases, outperforming various state-of-the-art methods. • The Graph Neural Network identifies the phase labels with accuracy higher than 90%. • Under noisy or unmetered demand, the Graph Neural Network achieves accuracy over 90%. • Under various Photovoltaic cases, the Graph Neural Network yields accuracy over 90%. • The Graph Neural Network yields accuracy of 92% in case of multi-phase connections. • The grid losses and uncertain assets’ parameters have minor impact on model accuracy.

Graph Neural Networks (GNNs) are the current state-of-the-art models in learning node representations for many predictive tasks on graphs. Typically, GNNs reuses the same set of model parameters across all nodes in the graph to improve the training efficiency and exploit the translationally-invariant properties in many datasets. However, the parameter sharing scheme prevents GNNs from distinguishing two nodes that are isomorphic and that the translation invariance property may not exhibit in real-world graphs. In this paper, we present Graph Representation Learning with Adaptive Mixtures (GRAM), a novel approach for learning node representations in a graph by introducing multiple independent GNN models and a trainable mixture distribution for each node. GRAM is a general framework that can be readily embedded into existing models for enhancing representation learning. It allows for adaptive node clustering with mixtures and joint optimization of all GNN models simultaneously. To achieve this, we develop an Expectation-Maximization algorithm with stochastic gradient descent in a scalable manner. Specifically, in the E-Step, we approximate the posterior probability distribution of the latent cluster membership based on the k-th independent GNN models in the last iteration. In the M-step, we update all GNN models and the mixture parameters based on the refined latent cluster membership for downstream tasks, e.g., node classification. To fully exploit the graph structural information, we further implement a regularization term on unlabeled nodes, which encourages adjacent nodes to have the same mixture distribution. By jointly optimizing the model in an end-to-end manner, GRAM learns to capture different abstract patterns on graphs for different node clusters, which yields discriminative representation for each node. We evaluate GRAM on five benchmark datasets with extensive experiments. GRAM is demonstrated to consistently boost state-of-the-art GNN variants in node classification tasks.

Traditional Graph Neural Network (GNN) learning patterns can only achieve optimal performance under the assumption of independent and identically distributed data. However, in real-world scenarios, numerous unpredictable factors give rise to the Out-of-Distribution (OOD) problem. Moreover, this distribution discrepancy may lead to unreliable predictions by GNNs models within unknown OOD domains. Therefore, this paper proposes a machine learning method to enhance the stability of GNN cross-site classification from the perspective of stable learning. The aim is to extract genuine causal features while eliminating spurious causal features. By introducing a feature sample weighting decorrelation technique in the random Fourier transform space and combining it with a baseline GNN model, a Stable-GNN model and a constrained sampling weight gradient update algorithm are designed. Its theoretical proof indicates that this algorithm can ensure the decrease in the loss, thus remedying the shortcomings of the original algorithm in weight updating. While guaranteeing predictive performance on the training distribution data, it also reduces prediction bias on data from unseen test distributions. Compared with the GNN model and the Stable-GNN model (S-GNN), the experimental results of the S-GNN model demonstrate that it not only surpasses the current state-of-the-art GNN models, but also offers a flexible framework for strengthening existing GNNs.

Developing fast and accurate computational models to simulate intricate physical phenomena has been a persistent research challenge. Recent studies have demonstrated remarkable capabilities in predicting various physical outcomes through machine learning-assisted approaches. However, it remains challenging to generalize current methods, usually crafted for a specific problem, to other more complex or broader scenarios. To address this challenge, we developed graph neural network (GNN) models with enhanced generalizability derived from the distinct GNN architecture and neural operator techniques. As a proof of concept, we employ our GNN models to predict finite element (FE) simulation results for three-dimensional solid mechanics problems with varying boundary conditions. Results show that our GNN model achieves accurate and robust performance in predicting the stress and deformation profiles of structures compared with FE simulations. Furthermore, the neural operator embedded GNN approach enables learning and predicting various solid mechanics problems in a generalizable fashion, making it a promising approach for surrogate modeling.

Graph neural networks (GNNs) have been used to tackle the few-shot learning (FSL) problem and shown great potentials under the transductive setting. However under the inductive setting, existing GNN based methods are less competitive. This is because they use an instance GNN as a label propagation/classification module, which is jointly meta-learned with a feature embedding network. This design is problematic because the classifier needs to adapt quickly to new tasks while the embedding does not. To overcome this problem, in this paper we propose a novel hybrid GNN (HGNN) model consisting of two GNNs, an instance GNN and a prototype GNN. Instead of label propagation, they act as feature embedding adaptation modules for quick adaptation of the meta-learned feature embedding to new tasks. Importantly they are designed to deal with a fundamental yet often neglected challenge in FSL, that is, with only a handful of shots per class, any few-shot classifier would be sensitive to badly sampled shots which are either outliers or can cause inter-class distribution overlapping. Extensive experiments show that our HGNN obtains new state-of-the-art on three FSL benchmarks. The code and models are available at https://github.com/TianyuanYu/HGNN.