With the advancement of big data technologies，interconnected entities in the real world generate massive amounts of data. As a data structure that efficiently and abstractly models entities and their relationships，graphs have been widely applied across various fields, and graph data mining has gradually become an important research direction in data science. However，real⁃world graph data not only contains rich relational information but is also inevitably subject to noise interference. Graph neural networks are particularly sensitive to adversarial perturbations，especially with regard to topological disturbances in graph structures，which can lead to significant drops in model performance. In this paper，we propose a Label⁃Guided text⁃attributed graphs representation learning Framework with Large Language Models (LGFLLM)，which is able to combine different graph neural network models and enhance their robustness for better node representations under topological perturbations. Specifically，on the textual attribute graph，the external knowledge base of the Large Language Models is used to generate predictive labels for each node in the graph，which is then combined with the node attribute to remove noisy edges in the graph and construct auxiliary edges to obtain a more reliable graph structure in noisy scenarios. Finally，more robust node representations are learned by combing with the base representation learning model. Experimental results on three citation datasets containing textual attributes show that LGFLLM performs better on the task of node classification in noisy scenarios than state⁃of⁃the⁃art methods.
Query rewriting has emerged as an essential technique in the Retrieval⁃Augmented Generation (RAG) paradigm，achieving remarkable success in enhancing the performance of knowledge⁃intensive tasks such as open⁃domain question answering (ODQA)，fact verification，and multi⁃hop reasoning. Traditional RAG pipelines typically follow a multi⁃stage workflow consisting of original query → rewritten query → retrieval → answer generation. In this conventional design，query rewriting，retrieval，and answer generation are handled by separate modules，such as a dedicated query rewriter and a downstream answer generation model. While this modular structure offers flexibility，it also leads to inefficiencies in inference，difficulty in joint optimization，and potential error propagation across stages.In this paper，we propose a novel unified query rewriting framework driven by reinforcement learning (RL) that integrates query rewriting，retrieval decision⁃making，and answer generation into a single large language model (LLM). Our approach eliminates the need for independent query rewriting and answer generation modules by enabling the LLM to directly perform end⁃to⁃end reasoning and generation within one inference process. Specifically，the model first produces a rewritten query that aims to maximize downstream task performance，triggers document retrieval，and incorporates the retrieved evidence into its working context. The same LLM then conducts reasoning over the combined query⁃evidence context and generates the final answer. By coupling all components into a single agent⁃like system，our method enhances reasoning efficiency，improves adaptability across tasks，and allows retrieval and generation strategies to be jointly optimized under a unified objective.To optimize this unified framework，we employ reinforcement learning with task⁃specific rewards that connect the quality of the rewritten query directly to the quality of the final answer. This feedback loop enables the LLM to autonomously learn optimal trade⁃offs between query reformulation，retrieval precision，and reasoning accuracy. Unlike supervised fine⁃tuning，our RL⁃based optimization allows the model to explore diverse rewriting strategies that may not appear in human⁃annotated training data but yield improved downstream performance. The experimental results show that，compared with the current strongest baseline，the tests using the EM standard and the LLM standard have improved by 4.82% and 5.00% respectively on the HotpotQA dataset，and by 11.50% and 6.98% respectively on the MuSiQue dataset. It has increased by 15.29% and 18.44% on the 2Wiki dataset.
To address the limitations of traditional gas production prediction models in capturing long⁃term dependencies and modeling nonlinear，multi⁃scale spatial interactions among wells，this study proposes a pretraining⁃enhanced spatio⁃temporal graph model for gas production prediction. In the pre⁃training phase，a Transformer⁃based time series masked autoencoder is employed to extract deep temporal features from oilfield production data，generating segment⁃level representations with global context awareness and mitigating the inadequacy of conventional models in temporal feature extraction. During the prediction phase，a spatio⁃temporal feature fusion mechanism effectively integrates these temporal representations with spatial dependencies among gas wells captured by a dynamic graph structure learning module，thereby overcoming the performance bottlenecks of traditional approaches in scenarios involving long⁃term memory deficiency and incomplete predefined graph structures. Experimental results demonstrate that the proposed model achieves a mean absolute error of 0.156，yielding error reductions of 6.6%，36.3%，26.9% and 67.9% compared to baseline models，thus delivering significant performance improvements. This research establishes a novel deep learning paradigm for spatio⁃temporal data modeling in petroleum exploration and represents a promising step toward the engineering application of large⁃scale spatio⁃temporal models in the energy sector.
Few⁃shot image classification remains a challenging task due to the scarcity of annotated samples. Metric learning has been widely adopted in this field，with the Euclidean distance commonly used to quantify the difference between query and support samples for category discrimination. However，the Euclidean distance is highly sensitive to outliers，which may result in misclassification. To address this limitation，this paper proposes a metric learning approach that leverages an optimized form of the Bray⁃Curtis distance as an alternative to the standard Euclidean distance. The Bray⁃Curtis distance is more robust against extreme values，but its original numerical range in few⁃shot learning is relatively narrow，limiting its ability to effectively capture inter⁃class differences. To overcome this drawback，we introduce a scaling factor to rescale the Bray⁃Curtis distance，thereby enhancing its applicability and discriminative capability in few⁃shot classification. Based on this idea，two network variants，ProtoNet_Bray and Meta DeepBDC_Bray，are constructed by integrating the optimized Bray⁃Curtis distance into ProtoNet and Meta DeepBDC，respectively. Extensive experiments on the MiniImageNet，TieredImageNet，and CUB⁃200⁃2011 datasets demonstrate that the proposed approach yields consistent performance gains over baseline methods.
Hypergraph neural networks (HNNs) have achieved remarkable success in processing complex data structures and capturing higher⁃order relationships among nodes. However，real⁃world hypergraph structures are typically sparse，meaning the hyperedges share few common nodes，which weakens the associations between the hyperedges and consequently limits the message⁃passing performance of HNNs. To address this issue，we propose a dual⁃view contrastive learning algorithm for sparse hypergraphs (DCSH). DCSH leverages rich node attribute information to construct an attribute hypergraph，supplementing the missing semantic similarity connections in the feature space of the structural hypergraph. It designs an attention⁃based structural hypergraph enhancement method to mitigate the sparsity of the structural hypergraph. Simultaneously，considering the semantic consistency between the hyperedges and the target nodes，it introduces an attention⁃based method to compute the influence of the hyperedges on node embeddings，generating node embeddings under different views. Finally，an adaptive fusion mechanism integrates multi⁃view node embeddings to determine the contribution of each view to the final node representation. To maintain structural consistency across views，DCSH employs contrastive losses between the enhanced structural hypergraph and both the original structural hypergraph and the attribute hypergraph to optimize the model. Experimental comparisons with nine classical algorithms on eight benchmark datasets demonstrate the effectiveness of the proposed approach.
Molecular property prediction is a fundamental task in various scientific domains，including drug discovery and material design. Given that molecular structures are naturally represented as graphs，numerous graph⁃based models have been developed to tackle this problem. However，as the molecular space continues to expand，these approaches face significant computational challenges，necessitating the development of lightweight models to enable faster and more efficient predictions. Despite this pressing need，effective solutions remain scarce. In this paper，we propose a novel two⁃level model lightweighting approach，named LW⁃MPP (Lightweighting Method for Efficient Molecular Property Prediction). First，we introduce a new knowledge distillation framework that converts large graph⁃based models into smaller SMILES (Simplified Molecular Input Line Entry System)⁃based models. Second，we apply a post⁃training pruning technique，which leverages masked search and reordering methods to further optimize model inference. Benchmark results on the large⁃scale PCQM4M⁃LSC (Predicting Quantum Mechanical Properties of Molecular Data⁃Large Scale Challenge) dataset demonstrate that our approach achieves a 3.82~17 times speedup in inference compared to traditional graph⁃based models，while maintaining near⁃optimal performance. Furthermore，our model outperforms most SMILES⁃Transformer⁃based models. When applied to specific downstream tasks with small⁃scale datasets from MoleculeNet，our model consistently achieves the best predictive accuracy in most cases.
Sarcasm detection aims to identify implicit sarcastic intent by uncovering semantic incongruity in text. Existing methods fail to fully exploit the syntactic dependencies between sentiment words and their corresponding targets，making it difficult for the models to capture semantic incongruity cues and leading to suboptimal performance. To address this issue，we propose a sarcasm detection model enhanced with dual⁃level syntactic dependency modeling，which strengthens the model's ability to capture syntactic relations at both the word and phrase levels. Specifically，at the lexical level，we construct a dependency syntax graph and a sentiment graph，using syntactic paths to associate sentiment words with their described objects，thereby revealing their global syntactic dependencies. At the phrase level，we identify nouns in the text and apply a dynamic weighting mechanism to construct phrases composed of nouns and their corresponding sentiment words，modeling their local dependencies. Finally，biaffine attention is employed to enable deep interaction between lexical⁃level and phrase⁃level information，enhancing the model's overall capacity for syntactic dependency modeling. Experiments conducted on the public IAC and Twitter datasets demonstrate the effectiveness of the proposed model.
Accurate prediction of stock market trends，including stock price movements，is crucial for economic growth and government macroeconomic management. It helps monitor and guide the stable operation of the stock market while reducing market risks. To address the significant non⁃stationarity and high noise characteristics of stock time series，as well as challenges such as data delays and missing values in practical applications，a meta⁃deep learning hybrid model，Meta⁃LSTM (Meta⁃Learning⁃Long Short⁃Term Memory)，is proposed. By leveraging prediction tasks based on similar stocks，the model achieves rapid generalization for target stock price prediction tasks through meta⁃learning，effectively addressing the problem of limited target stock data. Additionally，the model incorporates prediction errors and target stock features into the LSTM module，and optimizes the dynamic allocation of feature weights using the SE⁃MHA （Squeeze⁃and⁃Excitation Network⁃Multi⁃Head Attention） mechanisms. This approach enhances the model's ability to capture critical patterns in time series data. Experiments on the stock data from Shanghai Stock Exchange demonstrate that the Meta⁃LSTM model exhibits better stability and generalization，with prediction accuracy increasing by 5%~16% compared to other models.
Drone swarms can form flexible arrays to achieve agile beamforming，but their performance relies on precise node localization. Traditional swarm array position estimation is constrained by GPS accuracy，typically within 0.1~1 m. When the half⁃wavelength of the array calibration carrier is smaller than the positioning error，conventional array error calibration algorithms fail due to phase ambiguity. The challenge is further compounded when both communication time⁃delay errors and phase errors coexist in the array. To address these issues，this paper proposes a joint array error calibration algorithm based on a dual⁃frequency method. By constructing equivalent low⁃frequency measurements from dual⁃frequency differences and optimizing the frequency difference design，the algorithm overcomes the half⁃wavelength limitation in error modeling and achieves high⁃precision calibration. Simulation results demonstrate that the proposed method enables efficient joint error correction under various signal⁃to⁃noise ratio conditions，even when both positional errors and communication time⁃delay errors are present.
Firefly Algorithm (FA) is a heuristic algorithm used for solving optimization problems. However，FA has issues such as slow convergence speed in later stages and lower optimization accuracy，along with a lack of guidance in the optimization process. This paper proposes an improved Firefly Algorithm，X⁃FA，by integrating evolutionary factors and interpretability. X⁃FA introduces evolutionary factors to divide the entire optimization process into four stages，dynamically calculating the movement step size at different stages. This allows fireflies to search the entire search space more effectively in the initial optimization phase and converge more quickly to the optimal solution during the convergence phase. Additionally，the incorporation of interpretability methods provides guidance to the random movements of the fireflies. Finally，experimental analysis is conducted on six test functions，and the results indicate that X⁃FA achieves higher optimization accuracy and faster convergence speed.
Employing first⁃principles calculations via the Vienna Ab initio Simulation Package (VASP) based on density functional theory (DFT),this study investigates the structural stability,electronic properties,ferroelectric polarization,and piezoelectric characteristics of wurtzite⁃derived oxides β⁃AMO2(A=Li，Na; M=B，Al，Ga，In). The computational results demonstrate that the lattice parameters of the β⁃AMO2system exhibit regular expansion with increasing ionic radius of M3+. Due to the smaller ionic radius of Li⁺,Li⁃based materials generally exhibit lower
This study addresses the optimization of reference sensor configuration for active road noise control in vehicles. Based on a plate⁃cavity coupling model，the impact of reference sensor configuration on the noise reduction performance of active control systems is systematically investigated. Firstly，a numerical plate⁃cavity coupling model is established to obtain the frequency⁃domain transfer functions from the force sources on the plate to the reference and error points of the plate/cavity. The model is then used to analyze the coherence between reference and error signals and the system causality. Finally，experiments are carried out in an anechoic chamber. The results show that employing a number of reference sensors equal to the number of excitation sources achieves near⁃unity coherence across the full frequency band (except at spectral troughs). For system causality，simulation results based on the COMSOL finite element model demonstrate that acoustic reflections within the cavity significantly degrade the system causality of acoustic reference signals，resulting in substantially lower Wiener filter⁃based noise attenuation than that of vibration reference signals. Increasing the wall absorption coefficient effectively suppresses reflected sound and improves the noise reduction with acoustic reference signals to be comparable to that with vibration reference signals. The experimental results validate the simulation conclusions. This research provides a theoretical foundation for sensor placement in active road noise control systems for automotive applications.
Insulating sound transmission in low⁃frequency bands without blocking the airflow in a pipe remains one of the research focuses in the field of noise control. In particular，the pipe built⁃in sound insulation structure design has important value for theoretical research and practical application. Therefore，our work designs a sound barrier based on the Helmholtz resonators combined with the genetic algorithm. First，the ability of different geometries of necks to block acoustic waves is compared under the condition of identical surface areas. Then，the transmission loss effects of series，parallel，and linearly arranged resonant cavities are analyzed. The acoustic performance of various resonator⁃arranged methods is evaluated through theory and simulation，leading to the determination of the form for the final design architecture. Finally，the parameters of the parallel resonators are jointly optimized through the genetic algorithm. By accurately designing the geometric parameters of the neck and the cavity，two groups of parallel resonator arrays are finally designed as sound barriers. Under the condition of subwavelength size，the proposed sound barrier achieves a sound intensity transmission coefficient of 0.1 in the low⁃frequency range 314~1000 Hz while maintaining ventilation conditions.
This study systematically explored the dissolution laws of four typical carbonate rocks in Huanglong Formation，Qinglong Formation (mudstone，nodular limestone)，and Chuanshan Formation in the southern Jiangsu region through indoor simulation experiments. Based on regional geological surveys，representative rock samples were selected to conduct dissolution tests under different pH (2~5)，acidic solutions (sulfuric acid/hydrochloric acid)，and groundwater conditions (static/flowing)，analyzing the amount，rate，and main controlling factors of dissolution. The results showed that Qinglong Formation has the highest dissolution amount of mudstone (6.422 g at rest and 6.537 g at flow)，while Chuanshan Formation has the lowest dissolution amount of limestone (0.048 g at rest and 0.078 g at flow)，mainly controlled by mineral composition，structure，and specific surface area. At low pH (2~3)，sulfuric acid generally has better dissolution ability than hydrochloric acid，and the dissolution of Huanglong Formation limestone and Qinglong Formation marl is typical. The pattern is not significant when pH≥4. The amount of dissolution in the flowing state is significantly higher than that in the static state (especially at low pH)，because the water flow can remove surface sediment and maintain continuous dissolution; the dissolution rate increases with decreasing pH，and the flow state advances the dissolution transition point (slow to fast stage) by 12-48 hours; except for nodular limestone，all other rock samples exhibit a two⁃stage dissolution process of "slow to fast" at pH=2，with the transition time influenced by acid type and calcite content; due to its weak structure and significant fluctuations in dissolution rate，the nodular limestone of Qinglong Formation is prone to physical fragmentation under flowing conditions. This study reveals the main controlling mechanism of carbonate rock dissolution in the southern Jiangsu region，providing a theoretical basis for predicting regional karst development and preventing engineering disasters.
