Retrieval-augmented generation (RAG) is widely adopted for radiology report generation with medical vision-language models, leveraging external reports as linguistic references. However, existing RAG methods rely primarily on dense embedding similarity, which may retrieve reports that are semantically related yet clinically inconsistent with respect to presence or laterality constraints. Such inconsistencies are often propagated into generation, resulting in contradictory or unsupported findings. We propose an evidence-guided retrieval-augmented framework EviRAG that decomposes retrieval into structured and unstructured alignment levels. First, we induce structured clinical triplets from both query and database cases through targeted visual interrogation, projecting images into a shared evidence space. Triplet-level alignment enforces explicit agreement over presence and laterality variables, yielding a clinically admissible candidate set via structural ranking. Within this constrained space, we perform semantic alignment in a shared multimodal embedding space to capture nuanced descriptive correspondence. The top-ranked reports and query image are jointly fed into a medical vision-language model for report generation. Comprehensive experiments on radiology report generation benchmarks show that EviRAG substantially reduces clinical inconsistencies compared to strong medical vision-language baselines. The source code is available at https://github.com/liamgu06/EviRAG.

Graph-level classification involves analyzing the property of the whole graph, which is typically solved by using graph neural networks (GNNs). Existing efforts generally assume a balanced class distribution. However, real-world data often exhibit long-tailed distributions, i.e., tail classes have significantly fewer samples than head classes, and thus directly applying GNNs is eventually biased toward the head classes, resulting in limited generalization over the tail classes. Moreover, the predictions of existing algorithms are usually not trustworthy, and the trained classifiers remain ignorant to their predictive confidence. Towards this end, in this paper we develop a principled framework called GraphEVER for long-tailed graph-level classification. Technically, GraphEVER incorporates the beliefs of multiple experts and leverages the idea of subjective logic within the Dempster-Shafer Evidence Theory (DST). It can provide the evidence and uncertainty estimation for each expert, where the evidence is parameterized by a Dirichlet distribution to model class probability distribution, and the uncertainty is quantified via a well-defined theoretical framework. In this way, diverse experts can be integrated under DST to endow the classifier with both reliability and robustness. Moreover, we propose an evidence-based routing mechanism to dynamically assign experts, such that the tail classes can receive more attention, while the head classes can reduce redundant engaged experts, further cutting down the computational cost and improving the efficiency. Extensive experiments on seven datasets verify the superiority of our proposed framework.

Out-of-Distribution (OOD) detection, which seeks to identify samples deviating from the In-Distribution (ID) training distribution at test time, is crucial for building robust machine learning systems. While extensive efforts have been made for Euclidean data, OOD detection on graph-structured data remains relatively underexplored. On the one hand, the specific properties of a graph may be attributed to its substructures. On the other hand, acquiring labeled data for graph learning is typically time-consuming and labor-intensive. Toward this end, in this paper, we propose a novel kernel-enhanced graph substructure learning framework termed KEGOD for unsupervised graph OOD detection. Specifically, we introduce a learnable graph generator to construct the augmented graph view that preserves distinguishable structure information. Then, for both the input graph and augmented view, a graph neural network (GNN) branch and a graph kernel (GK) branch are incorporated to explore graph latent patterns. By performing multi-branch concordance learning on the extracted graph patterns, our KEGOD captures complementary ID structural semantics in both implicit and explicit manners, enabling reliable detection of OOD graphs through semantic inconsistency. Finally, we build a self-adaptive training mechanism to automatically control diverse sensitivities of the graph patterns. Experimental results on several public graph datasets reveal the superiority of our KEGOD. Our code is available at~ https://github.com/jamesyifan/KEGOD.

Graph neural networks (GNNs) have essentially taken over as the de facto model for learning graph-structured data. However, the majority of existing methods perform transductive learning in a known graph, which is unable to tackle abundant in-the-wild unseen graphs with potential domain shifts. Even worse, these graphs, accompanied by domain shifts on structural topology and node attributes, bring in vulnerable data bias and thus a huge drop in performance. To tackle this, we propose a novel GNN method named spectral optimal transport (SPOT) for effective domain generalization on graphs. Our method is motivated by the fact that the high-frequency graph spectrum is more likely to indicate domain differences. In particular, we formulate the structural augmentation as an optimal transport problem to retain low-frequency key knowledge and solve the problem using Sinkhorn-Knopp algorithm. In addition, we incorporate an adaptive perturbation strategy to deep features, where the direction of the additive noise is determined by the homophily degrees to maintain semantic properties. Accordingly, we meticulously construct a collection of real-world benchmark datasets to assess the domain generalization capability of our model on graphs, and extensive experiments confirm the effectiveness of our proposed SPOT.

Graph domain adaptation has emerged as a critical challenge in real-world applications, where labeled graph data is often scarce and expensive to obtain. While existing methods have shown promise, they typically require access to source domain data, which may be restricted due to privacy concerns or data regulations. To address these limitations, we investigate the challenging yet practical problem of source-free graph domain adaptation. We propose a new approach namedRobust CrossSupervision with Target Mining (ROSE) for this problem. ROSE achieves robustness by considering the complementary topology of graphs. The model consists of a message-passing branch for local semantic learning and a graph-kernel branch for global structural capture. Both branches are incorporated into a unified cross-supervision framework. To improve the robustness of the optimization process, we explore the context of the target domain, and divide the target data into discriminant set and anchor set. Then we incorporate the two tasks into a meta-learning optimization framework. Extensive experiments on benchmark datasets have demonstrated that our ROSE, compared with a wide range of baselines, always yields superior performance. The source code is available athttps://github.com/luo-junyu/ROSE.

Graph retrieval (GR), a ranking procedure that aims to sort the graphs in a database by their relevance to a query graph in decreasing order, has wide applications across diverse domains, such as visual object detection and myreddrug discovery. Existing Graph Retrieval (GR) approaches usually compare graph pairs at a detailed level and generate quadratic similarity scores. In realistic scenarios, conducting quadratic fine-grained comparisons is costly. However, coarse-grained comparisons would result in performance loss. Moreover, label scarcity in real-world data brings extra challenges. To tackle these issues, we investigate a more realistic GR problem, namely, efficient graph retrieval (EGR). Our key intuition is that, since there are numerous underutilized unlabeled pairs in realistic scenarios, by leveraging the additional information they provide, we can achieve speed-up while simplifying the model without sacrificing performance. Following our intuition, we propose an efficient model called Dual-TowerModel withDividing, Contrasting andAlignment (TowerDNA). TowerDNA utilizes a GNN-based dual-tower model as a backbone to quickly compare graph pairs in a coarse-grained manner. In addition, to effectively utilize unlabeled pairs, TowerDNA first identifies confident pairs from unlabeled pairs to expand labeled datasets. It then learns from remaining unconfident pairs via graph contrastive learning with geometric correspondence. To integrate all semantics with reduced biases, TowerDNA generates prototypes using labeled pairs, which are aligned within both confident and unconfident pairs. Extensive experiments on diverse realistic datasets demonstrate that TowerDNA achieves comparable performance to fine-grained methods while providing a 10× speed-up.

This paper investigates the problem of learning mesh-based physical simulations, which is a crucial task with applications in fluid mechanics and aerodynamics. Recent works typically utilize graph neural networks (GNNs) to produce next-time states on irregular meshes by modeling interacting dynamics, and then adopt iterative rollouts for the whole trajectories. However, these methods cannot achieve satisfactory performance in long-term predictions due to the failure of capturing long-term dependency and potential error accumulations. To tackle this, we introduce a new future-to-present learning perspective, and further develop a simple yet effective approach named Foresight And Interpolation (FAIR) for long-term mesh-based simulations. The main idea of our FAIR is to first learn a graph ODE model for coarse long-term predictions and then refine short-term predictions via interpolation. Specifically, FAIR employs a continuous graph ODE model that incorporates past states into the evolution of interacting node representations, which is capable of learning coarse long-term trajectories under a multi-task learning framework. Then, we leverage a channel aggregation strategy to summarize the trajectories for refined short-term predictions, which can be illustrated using an interpolation process. Through pyramid-like alternative propagation between the foresight step and refinement step, our proposed framework FAIR can generate accurate long-term trajectories, achieving a significant error reduction compared with the best baseline on four benchmark datasets. Extensive ablation studies and visualization further validate the superiority of our proposed FAIR.

Graph neural networks (GNNs) have shown remarkable success in graph-level classification tasks. However, most of the existing GNN-based studies are based on balanced datasets, while many real-world datasets exhibit long-tailed distributions. In such datasets, the tail classes receive limited attention during training, leading to prediction bias and degraded performance. To address this issue, a range of long-tailed learning strategies have been proposed, such as data re-balancing and transfer learning. However, these approaches encounter several challenges, including insufficient representation capacity for tail classes and their evaluation solely on uniform test data, limiting their capacity to handle unknown class distributions. To tackle these challenges, we introduce a novel framework, namely Knowledge-diverse Experts (KDEX) for long-tailed graph classification. Our KDEX leverages a dynamic memory module to enable the transfer of knowledge from head to tail, which improves the representation ability of the tail. To deal with unknown test distributions, KDEX introduces a knowledge-diverse expert training approach to train experts with different capacities in managing various test distributions. Moreover, we train the hierarchical router in a self-supervised manner to dynamically aggregate each knowledge-diverse expert during testing. Experimental results on multiple benchmarks reveal that our KDEX outperforms current baselines in both standard and test-agnostic long-tailed graph classification.

Collaborative Filtering (CF) is a pivotal research area in recommender systems that capitalizes on collaborative similarities between users and items to provide personalized recommendations. With the remarkable achievements of node embedding-based Graph Neural Networks (GNNs), we explore the upper bounds of expressiveness inherent to embedding-based methodologies and tackle the challenges by reframing the CF task as a graph-signal processing problem. To this end, we propose PolyCF, a flexible graph signal filter that leverages polynomial graph filters to process interaction signals. PolyCF exhibits the capability to capture spectral features across multiple eigenspaces through a series of Generalized Gram filters and is able to approximate the optimal polynomial response function for recovering missing interactions. A graph optimization objective and a pairwise ranking objective are jointly used to optimize the parameters of the convolution kernel. Experiments on three widely adopted datasets demonstrate the superiority of PolyCF over the state-of-the-art CF methods.

Graph neural networks (GNNs) are widely utilized to capture the information spreading patterns in graphs. While remarkable performance has been achieved, there is a new trending topic of evaluating node influence. We propose a new method of evaluating node influence, which measures the prediction change of a trained GNN model caused by removing a node. A real-world application is, "In the task of predicting Twitter accounts' polarity, had a particular account been removed, how would others' polarity change?". We use the GNN as a surrogate model whose prediction could simulate the change of nodes or edges caused by node removal. Our target is to obtain the influence score for every node, and a straightforward way is to alternately remove every node and apply the trained GNN on the modified graph to generate new predictions. It is reliable but time-consuming, so we need an efficient method. The related lines of work, such as graph adversarial attack and counterfactual explanation, cannot directly satisfy our needs, since their problem settings are different. We propose an efficient, intuitive, and effective method, NOde-Removal-based fAst GNN inference (NORA), which uses the gradient information to approximate the node-removal influence. It only costs one forward propagation and one backpropagation to approximate the influence score for all nodes. Extensive experiments on six datasets and six GNN models verify the effectiveness of NORA. Our code is available at https://github.com/weikai-li/NORA.git.

We aim at solving the problem of predicting people's ideology, or political tendency. We estimate it by using Twitter data, and formalize it as a classification problem. Ideology-detection has long been a challenging yet important problem. Certain groups, such as the policy makers, rely on it to make wise decisions. Back in the old days when labor-intensive survey-studies were needed to collect public opinions, analyzing ordinary citizens' political tendencies was uneasy. The rise of social medias, such as Twitter, has enabled us to gather ordinary citizen's data easily. However, the incompleteness of the labels and the features in social network datasets is tricky, not to mention the enormous data size and the heterogeneousity. The data differ dramatically from many commonly-used datasets, thus brings unique challenges. In our work, first we built our own datasets from Twitter. Next, we proposed TIMME, a multi-task multi-relational embedding model, that works efficiently on sparsely-labeled heterogeneous real-world dataset. It could also handle the incompleteness of the input features. Experimental results showed that TIMME is overall better than the state-of-the-art models for ideology detection on Twitter. Our findings include: links can lead to good classification outcomes without text; conservative voice is under-represented on Twitter; follow is the most important relation to predict ideology; retweet and mention enhance a higher chance of like, etc. Last but not least, TIMME could be extended to other datasets and tasks in theory.

Click-through rate (CTR) prediction, which aims to predict the probability of a user clicking on an ad or an item, is critical to many online applications such as online advertising and recommender systems. The problem is very challenging since (1) the input features (e.g., the user id, user age, item id, item category) are usually sparse and high-dimensional, and (2) an effective prediction relies on high-order combinatorial features (a.k.a. cross features), which are very time-consuming to hand-craft by domain experts and are impossible to be enumerated. Therefore, there have been efforts in finding low-dimensional representations of the sparse and high-dimensional raw features and their meaningful combinations. In this paper, we propose an effective and efficient method called the AutoInt to automatically learn the high-order feature interactions of input features. Our proposed algorithm is very general, which can be applied to both numerical and categorical input features. Specifically, we map both the numerical and categorical features into the same low-dimensional space. Afterwards, a multi-head self-attentive neural network with residual connections is proposed to explicitly model the feature interactions in the low-dimensional space. With different layers of the multi-head self-attentive neural networks, different orders of feature combinations of input features can be modeled. The whole model can be efficiently fit on large-scale raw data in an end-to-end fashion. Experimental results on four real-world datasets show that our proposed approach not only outperforms existing state-of-the-art approaches for prediction but also offers good explainability. Code is available at: \urlhttps://github.com/DeepGraphLearning/RecommenderSystems.

Online communities such as Facebook and Twitter are enormously popular and have become an essential part of the daily life of many of their users. Through these platforms, users can discover and create information that others will then consume. In that context, recommending relevant information to users becomes critical for viability. However, recommendation in online communities is a challenging problem: 1) users' interests are dynamic, and 2) users are influenced by their friends. Moreover, the influencers may be context-dependent. That is, different friends may be relied upon for different topics. Modeling both signals is therefore essential for recommendations. We propose a recommender system for online communities based on a dynamic-graph-attention neural network. We model dynamic user behaviors with a recurrent neural network, and context-dependent social influence with a graph-attention neural network, which dynamically infers the influencers based on users' current interests. The whole model can be efficiently fit on large-scale data. Experimental results on several real-world data sets demonstrate the effectiveness of our proposed approach over several competitive baselines including state-of-the-art models.