Recent advancements in Large Language Models (LLMs) and the proliferation of Text-Attributed Graphs (TAGs) across various domains have positioned LLM-enhanced TAG learning as a critical research area. However, the field faces significant challenges: (1) the absence of a unified framework to systematize the diverse optimization perspectives, and (2) the lack of a robust method capable of handling real-world TAGs, which often suffer from text and edge sparsity, leading to suboptimal performance. To address these challenges, we propose UltraTAG, a unified pipeline for LLM-enhanced TAG learning. UltraTAG provides a comprehensive and domain-adaptive framework that not only organizes existing methodologies but also paves the way for future advancements. Building on this framework, we propose UltraTAG-S, a robust instantiation designed to tackle sparsity issues in real-world TAGs. UltraTAG-S employs LLM-based text propagation and augmentation to mitigate text sparsity, while leveraging LLM-augmented node selection based on PageRank and edge reconfiguration strategies to address edge sparsity. Our experiments demonstrate UltraTAG-S significantly outperforms existing baselines, achieving improvements of 2.12% and 17.47% in ideal and sparse settings, respectively. Moreover, as the data sparsity ratio increases, the performance improvement of UltraTAG-S also rises.

Graph Neural Networks (GNNs) have become a pivotal framework for modeling graph-structured data, enabling a wide range of applications from social network analysis to molecular chemistry. By integrating large language models (LLMs), text-attributed graphs (TAGs) enhance node representations with rich textual semantics, significantly boosting the expressive power of graph-based learning. However, this synergy introduces critical vulnerabilities in both topology and text. Although specialized attack methods have been designed for each of these aspects, no work has yet unified them into a comprehensive approach. In this work, we propose the Interpretable Multi-Dimensional Graph Attack (IMDGA), a human-centric framework orchestrating multi-level perturbations across graph structure and textual features. IMDGA utilizes three tightly integrated modules to craft attacks that balance interpretability and impact, enabling a deeper understanding of Graph-LLM vulnerabilities. Through rigorous theoretical analysis and comprehensive empirical evaluations on diverse datasets and architectures, IMDGA demonstrates superior interpretability, attack effectiveness, stealthiness, and robustness compared to existing methods. By exposing these underexplored semantic vulnerabilities, our work offers valuable insights for improving Graph-LLM resilience. Our code is available at https://github.com/bwfan-bit/IMDGA.

Recently, graph neural network (GNN) has emerged as a powerful representation learning tool for graph-structured data. However, most approaches are tailored for undirected graphs, neglecting the abundant information in the edges of directed graphs (digraphs). In fact, digraphs are widely applied in the real world and confirmed to address heterophily challenges. Despite recent advancements, existing spatial- and spectral-based DiGNNs have limitations due to their complex learning mechanisms and reliance on high-quality topology, resulting in low efficiency and unstable performance. To address these issues, we propose Directed Random Walk (DiRW), a plug-and-play strategy for most spatial-based DiGNNs and also an innovative model which offers a new digraph learning paradigm. Specifically, it utilizes a direction-aware path sampler optimized from the perspectives of walk probability, length, and number in a weight-free manner by considering node profiles and topologies. Building upon this, DiRW incorporates a node-wise learnable path aggregator for generalized node representations. Extensive experiments on 9 datasets demonstrate that DiRW: (1) enhances most spatial-based methods as a plug-and-play strategy; (2) achieves SOTA performance as a new digraph learning paradigm. The source code and data are available at https://github.com/dhsiuu/DiRW.

Data compression is highly required to reduce the massive size of data while achieving lossless information over the transmission. In this paper, a novel multi-channel time series codec framework (LSCodec) is proposed to decouple long-term trend and short-term fluctuation using manually guided preprocessing data. The proposed LSCodec contains an encoder-decoder architecture network integrated with a group residual vector quantizer. The input data is decoupled through two paths layer by layer. In one path, a LSTM model is introduced for long-term trend feature learning. Another path produces fluctuation signal by subtracting between the real-time series and the trend signal to learn its hidden representation. The output of both path are then quantized using two individual residual vector quantization. A joint reconstruction loss combining its trend and fluctuation loss is used to support the training process. A balancer is used to stabilize training gradient of reconstruction loss to avoid local optimal solution or unstable state. Our experimental results show that the compression rate can vary to different bite rate according to strides setting. For multi-channel time series, it can compress data into average 10% with an acceptable reconstruction result. By reducing part of coding index, it is able to reconstruct part of curve with its main distribution. LSCodec can achieve a relatively good result in downstream task for a dataset that contains high ratio of anomaly. The proposed method can restore data distribution without losing its abnormal part. Several comparative studies are performed on with or without manually guided. The result shows the effectiveness of data guiding strategy. Code and models are available at https://github.com/HaiweiZuo/LSCodec.

In recent years, the increasing computational and storage demands of deep steganalysis models have drawn attention to lightweight architectures. While pruning algorithms for image steganalysis networks have been proposed, they often do not apply to networks equipped with mobile inverted bottleneck (MBConv) structures, such as EfficientNet. In this paper, we propose a Squeeze-and-Excitation Attention-based Pruning framework for image steganalysis networks, named SEAP. The method adopts a block-wise structured pruning strategy guided by the SE channel attention mechanism, where unimportant channels within each MBConv block are identified based on SE attention values and soft masks. Since pruning is conducted independently within each MBConv block and the input/output dimensions of the block remain unchanged, potential pruning conflicts across blocks are effectively avoided. In addition, we propose a sparsity regularization mechanism that adaptively adjusts the regularization strength based on the network structure, helping to preserve detection performance. Extensive experimental results demonstrate that the pruned network retains only a small fraction of the original network’s parameters and computational costs while achieving performance comparable to the original unpruned networks.

Discovering dense subgraphs in a graph is a fundamental graph mining task, which has a wide range of applications in social networks, biology and visualization to name a few. Even the problem of computing most cohesive subgraphs is NP-hard (like clique, quasi-clique, k-densest subgraph), there exists a polynomial time algorithm for computing the k-core and k-truss. In this paper, we propose a novel dense subgraph model, $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ , which leverages on a new type of important edges based on the basis of k-core and k-truss. We investigate the structural properties of the $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ model. Compared to k-core and k-truss, $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ can significantly discover the interesting and important structural information out the scope of k-core and k-truss. We study two useful problems of $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ decomposition and $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ search. In particular, we develop a k-core-truss decomposition algorithm to find all $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ in a graph G by iteratively removing edges with the smallest $${\mathsf {degree}}$$ - $${\mathsf {support}}$$ . In addition, we offer a $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ search algorithm to identifying a particular $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ containing a given query node such that the core number k is the largest. Extensive experiments on several web-scale real-world datasets show the effectiveness and efficiency of $${\mathsf {k}}$$ - $${\mathsf {core}}$$ - $${\mathsf {truss}}$$ model and proposed algorithms.