Time series anomaly detection is crucial in fields such as industrial monitoring, financial risk management, and network security. Graph Neural Networks (GNNs) have demon strated strong capabilities in capturing multivariate dependencies. However, existing methods often fail to adequately account for the temporal proximity between adjacent time points and are susceptible to the influence of weak or noisy connections during graph-based representation learning. To address these challenges, we propose LaGraph, a novel framework that integrates GNNs with a mask-optimized attention mechanism. Specifically, LaGraph decomposes input sequences into stable and trend components using an Expert Decomposition Block. The trend component is processed via a Multi-layer Convolution Block, while the stable component is modeled with a Proximity enhanced Graph Convolutional Network that incorporates a Laplacian kernel to capture local temporal dependencies. Additionally, a Mask-optimized Multi-head Attention Block, based on the Straight-Through Estimator (STE), mitigates the negative effects of less informative edges, enhancing both representation quality and reconstruction performance. Extensive experiments on five real-world benchmark datasets demonstrate that La Graph consistently outperforms state-of-the-art methods, veri fying its effectiveness and superiority for time series anomaly detection. To promote reproducibility and support future research, we have publicly released the full implementation at https://github.com/hit-zsc/LaGraph.

Therapeutic peptides, such as antimicrobial peptides(AMPs) and anticancer peptides (ACPs), are highly selective, low-toxicity agents that hold strong clinical potential. Designing peptides with specific biological functions remains challenging due to the vast combinatorial sequence space and the difficulty of modeling functional specificity, despite increasing interest. While deep generative models have recently shown promise in peptide design, most of them lack mechanisms for controllable functional output and underutilize protein language models (PLMs), which capture rich sequence-level dependencies. In this work, we propose MPTPep, a multi-task, prompt-aware fine-tuning framework built upon the pretrained PLM ProGen2, enabling controllable and function-specific therapeutic peptide generation. We introduce symbolic hard prompt tokens to encode peptide activity types, allowing the model to learn explicit function-sequence mappings within a unified generative architecture. Moreover, by jointly training on both AMP and ACP datasets, we exploit their biological overlap to boost the model to transfer knowledge from the AMP-rich corpus to boost ACP task performance. Extensive experiments demonstrate that MPTPep consistently outperforms state-of-the-art peptide generation models across both AMP and ACP sequence generation tasks, achieving superior functional scores, better sequence stability, and greater sequence diversity.

Longitudinal gene expression profiling provides crucial insights into dynamic biological processes, yet predictive modeling is severely hampered by irregular sampling, missing values, and high dimensionality. Mainstream approaches, including recurrent and transformer-based architectures, often struggle with the sparsity and temporal misalignment inherent in multi-visit transcriptomic datasets. We introduce G-Shapelets, a novel framework that learns to predict gene expression dynamics by matching discriminative local temporal patterns. Instead of modeling entire irregular trajectories, G-Shapelets identify multiscale temporal motifs-or shapelets-that serve as robust and interpretable building blocks for prediction. This approach is naturally resilient to gaps and irregular time intervals because it operates on local subsequences. By learning a dictionary of these shapelets and measuring their distance to segments of an input gene's trajectory, our model constructs a feature representation that effectively captures dynamic behavior. Extensive experiments on the longitudinal viral immunization dataset (HR-VILAGE3K3M) demonstrate that G-Shapelets achieves$\mathbf{4. 8 \%}$lower mean squared error and statistically significant improvements over recurrent, convolutional, and transformer baselines. Furthermore, we show that the learned shapelets correspond to meaningful biological patterns, offering interpretability that is absent in blackbox models. This framework opens new avenues for accurate and interpretable modeling of dynamic transcriptomic systems.

Accurate power load forecasting in industrial parks is crucial for optimizing energy management and operational efficiency. Existing models struggle with industrial load series' complex, multi-target nature and the need to integrate diverse exogenous variables. This paper introduces the Temporal Structure-Preserving Transformer (TSPT), a novel architecture that addresses these challenges by decomposing multi-target series into univariate series, enabling parallel processing and integrating exogenous data. The TSPT model incorporates the Gated Feature Fusion (GFF), which learns to capture multiscale temporal patterns from each target sequence and exogenous factors by preserving the temporal structure of the series. This parallel processing and the structure-preserving transformations allow TSPT to effectively integrate domain-specific knowledge, such as weather, production, and efficiency data, enhancing its forecasting performance. Comprehensive experiments on a real-world industrial park dataset demonstrate TSPT's superiority over state-of-the-art methods in handling complex, multi-target forecasting tasks with integrated exogenous variables. The proposed approach offers a pathway for scalable and accurate load forecasting in industrial settings, improving energy management and operational decision-making.

Forecasting carbon futures prices is a challenging task due to the complex and dynamic factors influencing them. Accurate forecasting can aid carbon market participants in hedging and optimizing their trading strategies. In this paper, we propose a novel feature selection method based on importance measures, aimed at selecting the most relevant and informative features for forecasting carbon futures prices. Our method introduces Gaussian noise to the input features, calculates the importance scores of the features, and determines the optimal threshold value for feature selection. We train and test different forecasting models on both the original and noisy feature sets using a 5-fold cross-validation approach. The importance score of each feature is calculated based on the error difference between the original and noisy feature sets. The optimal threshold value is determined based on the minimum prediction error obtained by ranking the features. We combine our feature selection method with different models to forecast carbon futures prices. The experimental results demonstrate that our method can effectively select useful features, outperforming variance thresholding and analysis of variance in feature selection. Moreover, our feature selection approach improves the prediction accuracy of different models. Our method is also robust in enhancing prediction accuracy across different models, test sets, time periods, and Gaussian noise levels.

This paper introduces a novel feature selection method, called Feature Selection based on Importance Measures (FS-IM), to enhance the forecasting of crude oil returns. FS-IM innovatively combines active learning with the application of Gaussian noise to input features and selects the most relevant features using an optimal threshold value. The paper applies a ridge regression (RR) model based on FS-IM (FS-RR) to identify the factors that have important information for crude oil return forecasting. The paper compares FS-IM with other dimension reduction methods such as Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA), and Independent Component Analysis (ICA). The results show that FS-IM can significantly improve model accuracy, demonstrating its effectiveness in finding key features. Moreover, FS-IM is more stable and consistent than other dimension reduction methods in enhancing the prediction accuracy in different scenarios, indicating its superior capability in capturing complex relationships between input and output variables. Furthermore, this study compares FS-RR model with other 13 prediction models by conducting experiments using a series of evaluation metrics, different statistical tests, and different step-ahead predictions and training sets. The results confirm that the RR model based on FS-IM can consistently outperform other model in terms of predictive performance and economic value, proving its effectiveness and robustness. This study contributes to the literature on crude oil price forecasting by addressing the challenges of high-dimensional and complex data, and by providing a robust, practical tool for professionals in energy economics and finance.

Abstract The prediction model is a major component within public health cyber‐physical systems, which supports decisions on prevention and control of diseases. Hand, foot, and mouth disease (HFMD) is one of the most common global infectious diseases with the highest incidence rate. Previous HFMD prediction models are mainly based on the time series that counted in equal‐grained time intervals. However, there are details in the time series counted in fine‐grained time intervals. To benefit from both equal‐grained and fine‐grained data, we proposed a dual‐grained representation (DGR) model. The DGR first represents inputted data to temporal patterns. Then, the represented patterns are consolidated to generate predictions. Experimental comparisons of the short‐term prediction performance are figured out by using real outpatient collections in Xiamen, China.

Hand, foot, and mouth disease (HFMD) is an infection that is common in children under 5 years old. This disease is not a serious disease commonly, but it is one of the most widespread infectious diseases which can still be fatal. HFMD still poses a threat to the lives and health of children and adolescents. An effective prediction model would be very helpful to HFMD control and prevention. Several methods have been proposed to predict HFMD outpatient cases. These methods tend to utilize the connection between cases and exogenous data, but exogenous data is not always available. In this paper, a novel method combined time series composition and local fusion has been proposed. The Empirical Mode Decomposition (EMD) method is used to decompose HFMD outpatient time series. Linear local predictors are applied to processing input data. The predicted value is generated via fusing the output of local predictors. The evaluation of the proposed model is carried on a real dataset comparing with the state‐of‐the‐art methods. The results show that our model is more accurately compared with other baseline models. Thus, the model we proposed can be an effective method in the HFMD outpatient prediction mission.