Multi-view graph clustering (MVGC) has made great progress in analyzing the interaction patterns of complex networks. Existing methods leverage different graph filters to obtain high- and low-pass signals and implement multi-view fusion. However, these filter-based methods face a scalability issue, which results in insufficient representation discrimination. Besides, they lack view-specific semantics in multi-view fusion, leading to poor information fidelity. To address these limitations, we propose a graph clustering framework with scalable graph filters and view-specific semantic fusion (SGSF-GC). SGSF-GC designs a Beta-based scalable graph filter and cohesion-based fusion mechanism to capture and integrate high- and low-frequency signals. Then SGSF-GC employs class activation mapping to capture semantics of view-specific representations for multi-view fusion. Finally, SGSF-GC conducts KL-based graph clustering. Extensive experiments on five public datasets with eleven baselines verify the utility and superiority of SGSF-GC.

Graph-based fraud detection garners extensive attention from academics and industry. However, existing methods fail to exploit the inconsistent information in heterophilic connections, overlooking important semantic information. In this paper, we propose a novel fraud detection approach against heterophily inconsistencies (HIGNN), which distinguishes malicious and unconscious connections in a heterophilic network. HIGNN utilizes the generated pseudo-label preference as guidance to exploit the latent semantic inconsistencies between malicious and unconscious connections, enhancing representative fraudulent characteristics. Rigorous experiments are conducted on four real-world datasets to verify the effectiveness of HIGNN. Compared to thirteen state-of-the-art baselines, the results show average performance improvements of 0.5% in AUC and 1.36% in F1 on four datasets, and the maximum performance improvement in F1 on the Elliptic dataset is 2.06%, demonstrating the effectiveness and superiority of the proposed HIGNN.

The rapid expansion of social media has intensified the spread of rumors, increasing the need for effective rumor detection. Previous research employing Graph Neural Networks (GNNs) fails to resist noise from the intricate information sources and model uncertainty caused by the lack of distribution characteristics, even tually leading to deficiencies in the robustness of models. To address these challenges, this paper proposes a novel robust rumor detection framework (RRD-N). In this framework, we apply data augmentation strategies to generate diverse graph views and leverage contrastive learning pretraining to learn the representative charac teristics of rumors. In particular, we adopt a node sampler and training loss based on Information Bottleneck (IB) theory to improve resistance to noise. Furthermore, a fine-tuning task utilizing Bayesian networks based on variational inference is implemented to obtain the latent distribution, tackling the issue of prediction uncer tainty. We conduct extensive experiments on two public datasets, Twitter15 and Twitter16. The results show performance improvements of at least 1.72% and 1.52% in accuracy compared with twelve state-of-the-art baselines, demonstrating the effectiveness and superiority of the proposed RRD-N. Our code is available at https://github.com/shaieesss/RRD-N.

Graph autoencoders (GAEs), a class of generative self-supervised learning methods, have demonstrated great potential in recent years. Typically, GAEs employ an encoder to map the input graph into a latent representation and a decoder to reconstruct the graph by recovering its characteristics, such as node features or structural information. However, GAEs that rely on feature reconstruction often fail to recover the unique information that differs from neighboring nodes, leading to excessive feature smoothness between neighboring nodes and sub-optimal performance. To address this issue, we propose two complementary strategies applied during the encoding and decoding phases, respectively. At the decoding stage, we develop a simple yet effective approach to preserve the distinctiveness between neighbors in the raw graph. We conceptualize the encoder-decoder architecture of GAEs as a teacher-student framework, where we compute pairwise node dissimilarities in both the original and reconstructed graphs and enforce a Kullback-Leibler divergence constraint to transfer distinctiveness from the input to the output space. At the encoding stage, we introduce a discriminative constraint that encourages decorrelation among similar node pairs, implemented via a covariance-based regularization that jointly considers node and neighborhood embeddings. Based on these strategies, we present ClearGAE, a GAE capable of reconstructing graphs while preserving their essential distinctions. Extensive experiments on three types of graph tasks demonstrate the effectiveness of ClearGAE. Moreover, our strategies are model-agnostic and can be seamlessly integrated as plug-and-play modules into other GAE variants.

This paper addresses the numerical instability in time-varying quadratic programming caused by ill-conditioned matrices, which often leads to error amplification and significant deviation from the true optimum. We propose a timevarying proximal point algorithm that reformulates the illposed problem into a sequence of well-conditioned subproblems, thereby improving the Hessian condition number while enforcing temporal consistency along the solution trajectory. In the static-parameter case, the method reduces to a proximal point algorithm with linear convergence, for which we establish a refined convergence rate. In the time-varying setting, we further prove convergence of the resulting solution sequence. Numerical experiments demonstrate that, relative to conventional approaches, the proposed framework significantly suppresses solution oscillations in high-condition-number regimes, yielding substantial improvements in numerical stability and decision smoothness.

High-frequency trading (HFT) demands adaptive strategies to navigate volatile markets. Current cutting-edge discrete sub-agent frameworks struggle with rigid market condition allocations, limiting adaptability. We propose a hierarchical framework with an attention-based meta-agent for dynamic sub-agent coordination. By leveraging market embeddings and reinforcement learning, the meta-agent optimally adjusts responsibility weights, enabling adaptive action aggregation across market regimes. Experiments on historical second-level HFT data show that the proposed framework outperforms state-of-the-art baselines, achieving a 42.15% total return and a 4.19 Sharpe ratio. Ablation studies validate the contributions of the dynamic sub-agent assign mechanism and multi-head attention mechanism, highlighting the framework's ability to adapt to market transitions and deliver superior performance.

The global shift to low-carbon energy systems has significantly increased demand for critical raw materials like lithium, cobalt, nickel, rare earth elements, and copper. These materials are essential for renewable technologies and energy storage. However, their extraction and processing produce significant carbon emissions and face challenges from supply chain vulnerabilities and price volatility. This review examines the complex relationship between carbon pricing mechanisms—such as carbon markets and taxes—and raw material markets. It explores the strategic importance of these materials, recent policy developments, and the transmission of carbon pricing impacts through supply chains. The review also analyzes the systemic risks created by carbon pricing, including regulatory uncertainty, market volatility, and geopolitical tensions. We then discuss financial tools and corporate strategies for managing these risks, such as carbon-linked derivatives and supply chain diversification. Finally, this review identifies key challenges and suggests future research to improve the resilience and sustainability of raw material supply chains. Here, resilience is defined as the capacity to adapt to carbon pricing volatility, geopolitical disruptions, and regulatory shocks, while maintaining operations. The paper concludes that coordinated policies and flexible risk management are urgently needed to support a reliable and sustainable energy transition.

The complexity of financial data, characterized by its variability and low signal-to-noise ratio, necessitates advanced methods in quantitative investment that prioritize both performance and interpretability.Transitioning from early manual extraction to genetic programming, the most advanced approach in the alpha factor mining domain currently employs reinforcement learning to mine a set of combination factors with fixed weights. However, the performance of resultant alpha factors exhibits inconsistency, and the inflexibility of fixed factor weights proves insufficient in adapting to the dynamic nature of financial markets. To address this issue, this paper proposes a two-stage formulaic alpha generating framework AlphaForge, for alpha factor mining and factor combination. This framework employs a generative-predictive neural network to generate factors, leveraging the robust spatial exploration capabilities inherent in deep learning while concurrently preserving diversity. The combination model within the framework incorporates the temporal performance of factors for selection and dynamically adjusts the weights assigned to each component alpha factor. Experiments conducted on real-world datasets demonstrate that our proposed model outperforms contemporary benchmarks in formulaic alpha factor mining. Furthermore, our model exhibits a notable enhancement in portfolio returns within the realm of quantitative investment and real money investment.

Masked graph autoencoders (GAEs) have attracted significant attention in recent years. GAEs typically leverage graph neural networks to reconstruct topological properties and node features. However, existing feature-based GAEs face performance bottlenecks, particularly on hard-to-reconstruct nodes, due to their excessive reliance on local aggregation. To address this limitation, we propose a novel framework, Global-Similarity-Enhanced Graph Autoencoder (GSE-GAE). GSE-GAE adopts a knowledge distillation strategy within a self-supervised teacher-student architecture. Specifically, a teacher module integrates raw features and topology with long-range structural augmentations for hard nodes, while a representation alignment loss ensures effective transfer of global knowledge to the student model. Extensive experiments demonstrate the superiority of GSE-GAE, providing new insights into improving performance.

Time series anomaly detection holds notable importance for risk identification and fault detection across diverse application domains. Unsupervised learning methods have become popular because they have no requirement for labels. However, due to the challenges posed by the multiplicity of abnormal patterns, the sparsity of anomalies, and the growth of data scale and complexity, these methods often fail to capture robust and representative dependencies within the time series for identifying anomalies. To enhance the ability of models to capture normal patterns of time series and avoid the retrogression of modeling ability triggered by the dependencies on high-quality prior knowledge, we propose a differencing-based contrastive representation learning framework for time series anomaly detection (DConAD). Specifically, DConAD generates differential data to provide additional information about time series and utilizes transformer-based architecture to capture spatiotemporal dependencies, which enhances the robustness of unbiased representation learning ability. Furthermore, DConAD implements a novel KL divergence-based contrastive learning paradigm that only uses positive samples to avoid deviation from reconstruction and deploys the stop-gradient strategy to compel convergence. Extensive experiments on five public datasets show the superiority and effectiveness of DConAD compared with nine baselines. The code is available at https://github.com/shaieesss/DConAD.

Graph Neural Networks (GNNs) play a significant role and have been widely applied in fraud detection tasks, exhibiting substantial improvements in detection performance compared to conventional methodologies. However, within the intricate structure of fraud graphs, fraudsters usually camouflage themselves among a large number of benign entities. An effective solution to address the camouflage problem involves the incorporation of complex and abundant edge information. Nevertheless, existing GNN-based methods frequently neglect to integrate this crucial information into the message passing process, thereby limiting their efficacy. To address the above issues, this study proposes a novel Gated Edge-augmented Graph Neural Network(GE-GNN). Our approach begins with an edge-based feature augmentation mechanism that leverages both node and edge features within a single relation. Subsequently, we apply the augmented representation to the message passing process to update the node embeddings. Furthermore, we design a gate logistic to regulate the expression of augmented information. Finally, we integrate node features across different relations to obtain a comprehensive representation. Extensive experimental results on two real-world datasets demonstrate that the proposed method outperforms several state-of-the-art methods.

Temporal Knowledge Graphs (TKGs) capture the dynamic nature of real-world facts by incorporating temporal dimensions that reflect their evolving states. These variations add complexity to the task of knowledge graph completion. Introducing temporal granularity can make the representation of facts more precise. In this paper, we propose Learning Temporal Granularity with Quadruplet Networks (LTGQ), which addresses the inherent heterogeneity of TKGs by embedding entities, relations, and timestamps into distinct specialized spaces. This differentiation enables a finer-grained capture of semantic information across the temporal knowledge graph. Specifically, LTGQ incorporates triaffine transformations to model high-order interactions between the elements of quadruples, such as entities, relations, and timestamps, in TKGs. Simultaneously, it leverages Dynamic Convolutional Neural Networks (DCNNs) to extract representations of latent spaces across different temporal granularities. By achieving more robust alignment between facts and their respective temporal contexts, LTGQ effectively improves the accuracy of temporal knowledge graph completion. The proposed model was validated on five public datasets, demonstrating significant improvements in TKG completion tasks, thereby confirming the effectiveness of our approach.

Fraudulent activities have significantly increased across various domains, such as e-commerce, online review platforms, and social networks, making fraud detection a critical task. Spatial Graph Neural Networks (GNNs) have been successfully applied to fraud detection tasks due to their strong inductive learning capabilities. However, existing spatial GNN-based methods often enhance the graph structure by excluding heterophilic neighbors during message passing to align with the homophilic bias of GNNs. Unfortunately, this approach can disrupt the original graph topology and increase uncertainty in predictions. To address these limitations, this paper proposes a novel framework, Dual-channel Heterophilic Message Passing (DHMP), for fraud detection. DHMP leverages a heterophily separation module to divide the graph into homophilic and heterophilic subgraphs, mitigating the low-pass inductive bias of traditional GNNs. It then applies shared weights to capture signals at different frequencies independently and incorporates a customized sampling strategy for training. This allows nodes to adaptively balance the contributions of various signals based on their labels. Extensive experiments on three real-world datasets demonstrate that DHMP outperforms existing methods, highlighting the importance of separating signals with different frequencies for improved fraud detection. The code is available at https://github.com/shaieesss/DHMP.

Time series anomaly detection is crucial for maintaining stable systems. Existing methods face two main challenges. First, it is difficult to directly model the dependencies of diverse and complex patterns within the sequences. Second, many methods that optimize parameters using mean squared error struggle with noise in the time series, leading to performance deterioration. To address these challenges, we propose a transformer-based framework built on decomposition (TransDe) for multivariate time series anomaly detection. The key idea is to combine the strengths of time series decomposition and transformers to effectively learn the complex patterns in normal time series data. A multi-scale patch-based transformer architecture is proposed to exploit the representative dependencies of each decomposed component of the time series. Furthermore, a contrastive learn paradigm based on patch operation is proposed, which leverages KL divergence to align the positive pairs, namely the pure representations of normal patterns between different patch-level views. A novel asynchronous loss function with a stop-gradient strategy is further introduced to enhance the performance of TransDe effectively. It can avoid time-consuming and labor-intensive computation costs in the optimization process. Extensive experiments on five public datasets are conducted and TransDe shows superiority compared with twelve baselines in terms of F1 score. Our code is available at https://github.com/shaieesss/TransDe.

Temporal knowledge graph reasoning aims to predict missing entities at future time steps, and as a critical task, it has attracted widespread attention in recent years due to its impressive ability to capture historical correlations and forecast future events. Although existing approaches, such as graph learning and logic rules, have partially addressed this problem, they still face limitations in modeling the uncertainty of future events especially when predicting rare or unseen facts. To address these challenges, we propose a diffusion model based on a selective attention mechanism (DMSA) for temporal knowledge graph reasoning. In our method, the encoder incorporates selective attention to emphasize key information, while the diffusion module introduces noise to enhance the model's capability to predict unseen events. By integrating selective attention with the diffusion module, our model improves both its memory and its ability to predict future, unseen events. Experimental results on five public datasets demonstrate that our proposed model achieves state-of-the-art performance across multiple evaluation metrics.

Traditional Mean-Variance (MV) portfolio optimization methods face challenges due to the separation of prediction and decision-making processes, rendering decisions highly sensitive to uncertain predictions and resulting in suboptimal performance. In this paper we propose an end-to-end (E2E) framework that integrates Robust Optimization (RO) to address uncertainties in expected returns. Firstly, we use the LSTM neural network approach to unify the prediction and decision-making stages. Secondly, to integrate the RO method into the neural network, we transform it so that it could remain differentiable and we use the transformed RO method as the loss function of the neural network. We then train and solve the model based on the returns of more than 1,000 stocks of NASDAQ from 2015 to 2020, and compare the results with the traditional MV method and other existing methods. The results show the superiority of the new model we proposed in this paper.

PurposeThis study aims to develop a fuzzy risk assessment model tailored to the adoption of offsite construction (OSC) in China, addressing the complexities and uncertainties associated with risk management in the construction sector. The model evaluates key risk factors, integrates expert opinions, and employs fuzzy decision-making techniques to improve the decision-making process.Design/methodology/approachThe research combines Interval-Valued Intuitionistic Fuzzy (IVIF)-DEMATEL with the Choquet Integral and IF-TOPSIS to assess and rank adoption risk factors (ARFs) in OSC. The Delphi method is employed for identifying potential risks, while the fuzzy approach accommodates the imprecision and subjectivity inherent in expert opinions. The IVIF-DEMATEL and IF-TOPSIS methods are integrated to assess relationships between risks and prioritize them accordingly.FindingsThe study reveals key ARFs impacting OSC adoption in China and identifies interdependencies among these factors. The proposed model outperforms existing approaches by providing more reliable risk evaluations, particularly in managing supply chain, logistics, quality control, and regulatory compliance risks. The case study emphasizes the robustness and flexibility of the model.Originality/valueThis research contributes to the field by introducing a novel fuzzy risk assessment model for OSC adoption, emphasizing the integration of expert opinions and the interconnection of adoption risks. It provides a comprehensive framework for better decision-making and risk management in the construction industry, offering insights into overcoming barriers to OSC in China.

Supply chain networks are susceptible to various risks, and effectively identifying them is crucial for economic stability. Current risk identification methods, however, often fail to account for the intrinsic uncertainties in both the financial data (node features) and the complex web of business relationships (network structure). To address this, we propose the Conditional Hierarchical Variational Graph Autoencoder (CTVGAE), a novel framework that employs a variational inference framework to capture the inherent randomness in both a company's financial data and its supply chain connections. By integrating a classifier with a conditional learning strategy, our model enhances its discriminative power to more reliably identify risky entities. Extensive experiments on a real-world supply chain dataset demonstrate that CTVGAE significantly outperforms nine baseline models, achieving an 8.79% increase in accuracy and a 5.23% gain in F1-score. These findings highlight the critical importance of modeling uncertainty for improving financial risk identification in complex networks and offer a powerful tool for practical applications.

Enterprise credit rating is crucial for risk management in the complex business environment. However, existing models encounter challenges in adeptly handling heterogeneous relationships and efficiently addressing risk propagation. This study introduces a novel framework for enterprise credit rating, the Risk Contagion Graph Neural Network (RCGNN), designed to overcome these challenges. The framework comprises four distinct modules: enterprise feature embedding, risk contagion intra-aggregation, risk contagion inter-aggregation, and credit rating prediction. Through the integration of advanced graph neural networks and attention mechanisms, RCGNN effectively captures intricate risk propagation relationships among enterprises. Extensive experiments were conducted on a self-constructed dataset of Chinese listed companies, and the results demonstrate that RCGNN outperforms traditional machine learning methods and homogeneous graph neural network models in credit rating tasks.