The missing of graph attributes poses a significant challenge in graph representation learning. Some existing graph attribute completion methods adopt the shared-space hypothesis or employ end-to-end frameworks to perform single-attribute imputation. However, these models can only generate one single attribute with a few specific patterns that either adhere to prior knowledge or are optimal for downstream tasks, making it difficult to capture the full range of variations in the target attribute distribution. This limitation negatively impacts the model's generalizability and efficiency. Therefore, to address this issue, we proposed a new method based on a graph denoising diffusion model, called Multi-attribute Imputation Graph Denoising Diffusion Model (MIGDiff), which can generate multiple high-quality attributes. Specifically, it employs a Dual-source Auto-encoder on existing attributes and graph topology to extract reliable knowledge, which serves as a condition for training the diffusion module. Within diffusion, noise is added to the structural embeddings of nodes without attributes in the forward process. In the reverse process, a Structure-aware Denoising Network is devised to integrate feature and structural information via an attention mechanism and to perform neighbor-guided refinement based on graph connectivity, thereby enhancing denoising and accurately recovering missing attributes while effectively maintaining structural consistency and distributional fidelity. During generation, multiple initial values are sampled to produce diverse attribute imputations, avoiding focusing on a few easy-to-learn patterns. Extensive experiments conducted on four public datasets highlight the state-of-the-art performance of MIGDiff in both attribute imputation and node classification tasks.

Graph contrastive learning (GCL) aims to learn representations by bringing semantically similar graphs closer and pushing dissimilar ones farther apart without label supervision. Hard negatives, which refer to graphs that have different labels but similar embeddings to the target graph, play a key role in improving representation discrimination. However, current methods that generate both high-quality positives and hard negatives face two challenges: (1) Hard negative sample generation often suffers from class imbalance, resulting in unequal attention across classes and reduced discriminative power in the learned representations. (2) The typical binary positive sample generation approach, which divides the graph into important and unimportant semantic regions, overlooks regions that negatively impact semantics and mislead model predictions. To address these issues, we introduce a novel method named BalanceGCL, which enhance graph contrastive learning with balanced hard negatives and fine-grained semantic-aware positives. BalanceGCL comprises two modules: Balanced Hard Negative graphs generation (BHN) and Fine-grained Semantic-aware Positive graphs generation (FSP). Inspired by the counterfactual mechanism, BHN generates balanced hard negatives that remain structurally similar to the original graph while inducing a controlled semantic shift. To ensure class balance, BHN iteratively constructs one hard negative sample for each class, ensuring an even distribution of negative samples across all alternative categories. FSP leverages the semantic differences between original graphs and balanced hard negatives to identify positively contributing, negatively contributing, and unimportant regions. By enhancing the influence of positive contributors, suppressing negative ones, and perturbing unimportant areas, it generates more reliable and semantically complete positive samples. The proposed method outperforms state-of-the-art GCL techniques across 14 datasets in graph classification and transfer learning tasks, demonstrating its effectiveness in tackling class imbalance and identifying fine-grained semantic-aware regions.

Spatial transcriptomics technologies carry out advanced sequencing analysis of molecular profiles with a spatial context, providing multi-source information essential for elucidating biological regulatory mechanisms. Nonetheless, it poses challenges in the integration of raw spatial coordinates with high-dimensional gene expression profiles in their native feature space. While spatial-aware methods effectively aggregate molecular information from local spatial neighborhoods, they fail to explore the long-range relationships associated with gene expression data. To address this issue, this paper introduces a novel approach termed GraphSTAR that encodes both spatial and gene expression data into undirected graphs, characterizing the local spatial proximity and global transcriptional similarity, respectively. Through a graph aggregation process, GraphSTAR integrates these diverse data sources within a joint graph structure, effectively modeling both local neighborhood relationships and long-range functional associations. Subsequently, a reassembled graph neural network is established by incorporating the graph aggregation into the feed-forward propagation using proximal operators, progressively refining spatial-informed latent representation to decipher spatial expression patterns of genes. Extensive experiments on benchmark datasets demonstrate that GraphSTAR outperforms state-of-the-art methods in both spatial domain identification and cell-type annotation tasks.

Large Language Models (LLMs) excel at natural language processing tasks, but their massive size leads to high computational and storage demands. Recent works have sought to reduce their model size through layer-wise structured pruning. However, they tend to ignore retaining the capabilities in the pruned part. In this work, we re-examine structured pruning paradigms and uncover several key limitations: 1) notable performance degradation due to direct layer removal, 2) incompetent linear weighted layer aggregation, and 3) the lack of effective post-training recovery mechanisms. To address these limitations, we propose CoMe, including a progressive layer pruning framework with a Concatenation-based Merging technology and a hierarchical distillation post-training process. Specifically, we introduce a channel sensitivity metric that utilizes activation intensity and weight norms for fine-grained channel selection. Subsequently, we employ a concatenation-based layer merging method to fuse the most critical channels in the adjacent layers, enabling a progressive model size reduction. Finally, we propose a hierarchical distillation protocol, which leverages the correspondences between the original and pruned model layers established during pruning, enabling efficient knowledge transfer. Experiments on seven benchmarks show that CoMe achieves state-of-the-art performance; when pruning 30% of LLaMA-2-7b's parameters, the pruned model retains 83% of its original average accuracy.

Spatio-temporal time series forecasting has attracted great attention in various fields, including climate, power, and traffic forecasting. Recently, Spatio-temporal Graph Neural Networks (STGNNs) have shown promising performance in modeling spatial dependencies based on graph neural networks (GNNs) and temporal dependencies based on temporal learning modules. However, most STGNNs do not effectively integrate explicit and implicit relationships between nodes, nor do they adequately capture long and short-term time dependencies. To address these challenges, this paper presents a Quaternion Spatio-temporal Graph Neural Network(QSTGNN). Specifically, the quaternion spatio-temporal graph is constructed firstly, such that the information of both short and long-term time steps are preserved in quaternion feature tensor, and information of multiple explicit graphs and implicit graph are integrated in quaternion graph adjacency matrix. Then, two modules are designed: a 1D quaternion convolution module and a quaternion graph convolution module. In the 1D quaternion convolution module, complex temporal correlations among short and long-term time steps can be well exploited by 1D quaternion convolution operator based on the quaternion Hamilton product. In the quaternion graph convolution module, quaternion graph convolution is designed to characterize nonlinear dependencies among multiple spatial graphs, including explicit and implicit graphs. Extensive experiments are conducted on six datasets, and the results show that QSTGNN achieves state-of-the-art performances over the existing ten methods. Explainable analysis presents that multiple spatial correlations can accurately illustrate the traffic flow and road functional information in real traffic roads.

Multiview subspace clustering has shown promising performance in multimedia and data mining applications. However, its employment in large-scale datasets is limited due to its quadratic or even cubic computational complexity. The anchor graph strategy, which selects a few important samples (anchors) to represent the whole data for different views, has been introduced to address this challenge. These methods rely on a heuristic assumption that the correspondence and class structures between the sets of anchors across different views are the same. This assumption ignores the difference in the ordering of anchors with respect to their associated classes and the number of anchors belonging to the same class from different views. As a result, this can lead to unsatisfactory clustering results due to incorrect anchorwise and classwise alignments. To tackle this issue, this article proposes an anchor-based multiview subspace clustering with anchorwise and classwise alignments (AMCA2) method. Specifically, the proposed method simultaneously aligns and fuses multiple anchor graphs anchor wisely and class wisely via learning permutation matrices and utilizing the Hadamard product. To further enhance the clustering performance of AMCA2, we propose a novel anchor selection method called kernel anchor selection (KAS) to select more representative anchors. Extensive experiments on ten benchmark datasets are conducted to show the superiority and effectiveness of AMCA2over the state-of-the-art methods.

With the development of high-throughput technologies, a massive scale of pharmacological and genomic data has been accumulated, which enables the discovery of the correlation between oncogenic genes and therapeutic drugs. Generally, genes with similar functions tend to be related to similar drugs and vice versa. Previous methods detect such associations between gene modules and drug modules based on the similarity of individual genes and drugs, resulting in an inaccurate capture of module-level interactions. How to fully leverage the underlying modules within genes and drugs is a key challenge when identifying regulatory relationships between gene and drug modules. In this paper, we propose a module-level gene-drug interaction identification model via hierarchical optimal transport (H-OT). Particularly, prior knowledge of genes or drugs is integrated to uncover the underlying module of genes or drugs with similar biological functions. Moreover, the optimal associations between gene and drug modules are determined by minimizing high-level OT distance between them, the cost function specified in high-level OT is automatically learned by low-level OT, which incorporates module patterns within genes and drugs. Experiments conducted on synthetic datasets demonstrate that our model exhibits superior performance than six state-of-the-art methods. Additionally, our evaluation of real drug-gene data highlights the model’s statistical power. The gene-drug modules identified by our approach reveal closely related gene-drug interactions and significantly enrich pathways associated with cancer.

The field of multimedia research has witnessed significant interest in leveraging multimodal pretrained neural network models to perceive and represent the physical world. Among these models, vision-language pretraining (VLP) has emerged as a captivating topic. Currently, the prevalent approach in VLP involves supervising the training process with paired image-text data. However, limited efforts have been dedicated to exploring the extraction of essential linguistic knowledge, such as semantics and syntax, during VLP and understanding its impact on multimodal alignment. In response, our study aims to shed light on the influence of comprehensive linguistic knowledge encompassing semantic expression and syntactic structure on multimodal alignment. To achieve this, we introduce SNARE , a large-scale multimodal alignment probing benchmark designed specifically for the detection of vital linguistic components, including lexical, semantic, and syntax knowledge. SNARE offers four distinct tasks: Semantic Structure, Negation Logic, Attribute Ownership, and Relationship Composition. Leveraging SNARE , we conduct holistic analyses of six advanced VLP models (BLIP, CLIP, Flava, X-VLM, BLIP2, and GPT-4), along with human performance, revealing key characteristics of the VLP model: (i) Insensitivity to complex syntax structures, relying primarily on content words for sentence comprehension. (ii) Limited comprehension of sentence combinations and negations. (iii) Challenges in determining actions or spatial relations within visual information, as well as difficulties in verifying the correctness of ternary relationships. Based on these findings, we propose the following strategies to enhance multimodal alignment in VLP: (1) Utilize a large generative language model as the language backbone in VLP to facilitate the understanding of complex sentences. (2) Establish high-quality datasets that emphasize content words and employ simple syntax, such as short-distance semantic composition, to improve multimodal alignment. (3) Incorporate more fine-grained visual knowledge, such as spatial relationships, into pretraining objectives. 1

Multi-domain biological network association and clustering have attracted a lot of attention in biological data integration and understanding, which can provide a more global and accurate understanding of biological phenomenon. In many problems, different domains may have different cluster structures. Due to rapid growth of data collection from different sources, some domains may be strongly or weakly associated with the other domains. A key challenge is how to determine the degree of association among different domains, and to achieve accurate clustering results by data integration. In this paper, we propose an unsupervised learning approach for multi-domain network association by using block signed graph clustering. In particular, with consistency weights calculation, the proposed algorithm automatically identify domains relevant to each other strongly (or weakly) by assigning them larger (or smaller) weights. This approach not only significantly improve clustering accuracy but also understand multi-domain networks association. In each iteration of the proposed algorithm, we update consistency weights based on cluster structure of each domain, and then make use of different sets of eigenvectors to obtain different cluster structures in each domain. Experimental results on both synthetic data sets and real data sets (including neuron activity data and gene expression data) empirically demonstrate the effectiveness of the proposed algorithm in clustering performance and in domain association capability.