International Journal of Intelligent Systems最新文献

Current​ fine-tuning techniques for large pretrained language models (LLMs) face significant challenges, particularly regarding the high computational costs associated with adapting billions of parameters and their limitations in effectively addressing diverse language understanding tasks. These methods often result in an inability to manage inter-task dependencies effectively, leading to underutilization of inter-task information. To address these issues, we propose tasks-embedded reparameterization (TER), a novel parameter-efficient fine-tuning framework that exploits multitask learning to enhance task-specific capabilities. The TER model integrates prompt tuning and multitask reparameterization, merging task-specific experts and hidden states of target tasks in a unified model framework. Furthermore, it employs a dynamic, task-oriented gating mechanism to optimize the prompts output by the model. This method dynamically adjusts the parameters according to the differing requirements of the task, ensuring that the model optimally adjusts the parameters according to the specific requirements of the task, so that the task can find a suitable balance between different tasks and improve knowledge sharing and task adaptability. Experimental evaluations using the SuperGLUE benchmark demonstrate that TER consistently outperforms existing parameter-efficient fine-tuning techniques in both performance and computational efficiency, offering a promising solution for task-specific language understanding in both research and industry.

Detecting financial reporting fraud is vital for preserving market integrity and protecting investors from substantial losses. Yet, the challenges of high dimensionality and noisy financial data often undermine the effectiveness of existing financial fraud detection systems. To address these issues, this study proposes SISAE-METADES, a novel framework that integrates a supervised input-enhanced stacked autoencoder (SISAE) with a meta-learning–based dynamic ensemble selection (METADES) strategy. Unlike conventional stacked autoencoders, SISAE concatenates the original input at each encoding stage and incorporates label supervision, thereby learning task-relevant and class-discriminative representations. These enriched deep features improve both the diversity and competence of base classifiers and enable METADES to achieve more reliable local competence estimation. We validate the proposed framework using financial statement data from Chinese A-share listed companies (2005–2023), covering 71 indicators. Experimental results show that SISAE-METADES significantly outperforms standalone SISAE, traditional METADES, and several state-of-the-art baselines. In particular, it achieves substantial improvements in accuracy, recall, and F1-score, underscoring the robustness and effectiveness of combining supervised deep representation learning with dynamic ensemble selection for financial fraud detection. These findings highlight the framework’s practical significance in reducing investor losses, strengthening market confidence, and promoting the stability of the financial system.

With the exponential growth of multimedia content, visual sentiment classification has emerged as a significant research area. However, it poses unique challenges due to the complexity and subjective nature of the visual information. This can be attributed to the significant presence of semantically ambiguous images within the current benchmark datasets, which enhances the performance of sentiment analysis but ignores the differences between various annotators. Moreover, most current methods concentrate on improving local emotional representations that focus on object extraction procedures rather than utilizing robust features that can effectively indicate the relevance of objects within an image through color information. Motivated by these observations, this paper addresses the need for efficient algorithms for labeling and classifying sentiment from visual images by introducing a novel hybrid model, which combines content-based image retrieval (CBIR) and a multi-input convolutional neural network (CNN). The CBIR model extracts color features from all dataset images, creating a numerical representation for each. It compares a query image to dataset images’ features to find similar features. This process continues until the images are grouped according to color similarity, which allows accurate sentimental categories based on similar features and feelings. Then, a multi-input CNN model is utilized to extract and efficiently incorporate high-level contextual visual information. This model comprises 70 layers, with six branches, each containing 11 layers. It seeks to facilitate the fusion of complementary information by incorporating multiple input categories that differ according to the color features extracted by the CBIR technique. This feature enables the model to understand the target and generate more precise predictions fully. The proposed model demonstrates significant improvements over existing algorithms, as evidenced by evaluations of six benchmark datasets of varying sizes. Also, it outperforms the state of the art in sentiment classification accuracy, getting 87.88%, 84.62%, 84.1%, 83.7%, 80.7%, and 91.2% accuracy for the EmotionROI, ArtPhoto, Twitter I, Twitter II, Abstract, and FI datasets, respectively. Furthermore, the model is evaluated on two newly collected large datasets, which confirm its scalability and robustness in handling large-scale sentiment classification tasks, and thus achieves a significant accuracy of 85.21% and 83.72% with the BGETTY and Twitter datasets, respectively. This paper contributes to the advancement of visual sentiment classification by offering a comprehensive solution for analyzing sentiment from images and laying the foundation for further research.

In this paper, we introduced a novel methodology to build a classification-forecast model used for financial risk forewarning. For the first step, we utilize the K–S test, Mann–Whitney U test, and Pearson’s correlation to select variables. Then, we employ CRITIC and fuzzy comprehensive evaluation (FCE) methods to score the risk of listed companies. Following this, self-organizing maps (SOM) clustering is utilized to segment the samples into five distinct risk levels. For the second step, we utilized triangulated maximally filtered graph (TMFG) and maximally filtered clique forest (MFCF) to minimize the number of indicators based on the dependent relationships between variables. These are then combined with Gaussian Markov random field (GMRF) and Copula algorithms to address nonlinear situations, forming what we refer to as the LoGo model. To further enhance the accuracy of LoGo models, we utilize the square Mahalanobis distance to compute the log-likelihoods as part matrix. The results reveal that the enhanced LoGo model with part matrix improves average accuracy by 7% compared with the original models without part matrix, albeit with a tenfold increase in execution time. MFCF demonstrates superior performance over TMFG in linear situations, achieving a 40% higher accuracy. However, under nonlinear circumstances, TMFG only requires half the execution time of MFCF, yet achieves a slightly higher average accuracy. Furthermore, compared with the widely used CNN models, the enhanced LoGo models show superior performance as they achieved closed accuracy in a shorter time.

While Large Language Models (LLMs) demonstrate strong translation capabilities, optimizing their output towards human-level refinement necessitates reasoning-guided approaches that move beyond simple generation. This paper introduces Multidimensional Feedback and Postedit Thought (MFPE), a novel framework specifically designed for reasoning-guided LLM translation optimization. MFPE operationalizes this guidance by leveraging multidimensional postediting feedback, which acts as explicit reasoning signals to the LLM. This feedback mechanism simulates the human postediting process, where errors are systematically identified and corrected. Generated by a dedicated optimization model trained on a synthetic dataset (using GLM-4 and inspired by multidimensional quality metrics (MQM), this feedback provides fine-grained error details including spans, categories, and quantities from initial LLM translations. We conduct experiments across four language pairs: Chinese-English, German-English, English-Chinese, and English-German. The results show that fine-tuning with structured, reasoning-like feedback significantly enhances translation quality and outperforms standard bilingual fine-tuning approaches. Our findings highlight the effectiveness of simulating postediting reasoning through structured feedback, offering a promising direction for harnessing and improving the inferential capabilities of LLMs for complex tasks like high-quality machine translation.

This article presents a novel neural network–based approach for designing effective control policies for Caputo-type nonlinear fractional-order systems. The proposed approach iteratively refines the neural network to generate a control policy that stabilizes the system within a predefined neighborhood around the zero equilibrium. Stability of the controlled system is guaranteed by rigorously formulated theorems and empirically verified using a neural Lyapunov function. The effectiveness of the proposed methodology is demonstrated through simulations on two classical Caputo fractional-order systems, showcasing its capability to ensure stability and its potential applicability to a broader range of fractional-order nonlinear systems.

The slime mold algorithm (SMA) is a metaheuristic optimization algorithm that simulates the foraging behavior of slime molds. Compared to other optimization algorithms, SMA has fewer parameters, faster convergence speed, and stronger optimization capabilities. However, the standard SMA uses two randomly selected individuals to guide the search direction of the population, which results in excessive randomness during the search process. This can lead to the loss of valuable information and waste computational resources. To overcome these limitations, this study proposes an enhanced slime mold algorithm (S2SMA) based on a spiral sensing search mechanism. The main contributions of this study are as follows: Firstly, a fitness–distance balanced oscillation search mechanism is introduced to solve the issue of lack of guidance in the individual oscillatory search phase in the original SMA, thus enhancing the global exploration ability of the algorithm. Secondly, the spiral sensing search mechanism is introduced, reshaping the random redistribution behavior in SMA. This aims to fully utilize the effective information in the existing population, improve search efficiency, and enhance population diversity. Finally, the computational logic of SMA is restructured based on the existing parameters, improving the algorithm’s performance while avoiding additional computational overhead. To validate the effectiveness of the proposed S2SMA, experiments were conducted on 71 test instances from the IEEE CEC2017 and IEEE CEC2021 benchmark sets, as well as three engineering problems. The algorithm was compared with classical algorithms, high-performance algorithms, and advanced SMA variants. Experimental results show that S2SMA outperforms the classical algorithms, high-performance algorithms, and other SMA variants in terms of both performance and robustness, demonstrating its potential application in engineering optimization.

This study presents a novel hybrid quantum-classical framework for rice disease diagnosis, leveraging variational quantum circuits (VQCs) to address the limitations of traditional and deep learning models in precision agriculture. The proposed Quantum Variational Rice Disease Network (QVRDN) integrates quantum feature encoding, variational quantum processing, and adaptive optimization to achieve superior classification accuracy, efficiency, and robustness. Using a curated dataset of 3000 annotated rice leaf images spanning major disease categories, the QVRDN framework applies dimensionality reduction and quantum angle encoding to transform the image features into quantum states, which are then processed by parameterized quantum circuits for disease classification. Experimental results demonstrate that QVRDN outperforms classical models, including SVM, random forest, CNN, and ResNet50-achieving, the highest accuracy of 97.8%, faster inference times, and greater resilience to noise and limited data. The compact design of the framework enables edge deployment without GPU dependency, making it suitable for resource-constrained agricultural environments. By demonstrating the feasibility and advantages of quantum machine learning in crop health monitoring, this study establishes a foundation for quantum-enhanced, data-efficient agricultural diagnostics and paves the way for future advances in intelligent, field-ready quantum geoinformatics systems.

To address the limitations of traditional computational fluid dynamics (CFD) simulations, such as high computational cost, long processing times, and limited scalability, this study identifies the inefficiencies of existing data-driven prediction methods, which often lack spatial–temporal coordination mechanisms and fail to capture fine-grained dynamic features of UAV airflow fields. We propose a novel deep learning model, VAN-ConvLSTM, for rapid and accurate prediction of UAV downwash airflow. Unlike conventional ConvLSTM-based frameworks, which struggle with modeling long-range dependencies and detailed spatial variations, our model introduces a visual attention unit (VAN) to enhance spatiotemporal sensitivity. The model architecture combines a convolutional encoder for spatial feature extraction, a VAN module for attention-guided temporal modeling, and a ConvLSTM decoder for sequence generation. This synergistic design improves both the accuracy and interpretability of airflow prediction. Experimental results show that the VAN-ConvLSTM model achieves an SSIM score of 0.96, demonstrating high consistency with CFD simulations. Compared to baseline methods, our model reduces error while improving stability and spatial fidelity. Ablation studies further validate the individual contributions of VAN and ConvLSTM modules. The results, verified through three representative case studies, confirm that VAN-ConvLSTM outperforms state-of-the-art approaches across multiple evaluation metrics, while offering significantly enhanced computational efficiency. This demonstrates its strong potential as a reliable and scalable alternative to traditional CFD methods in rotor airflow prediction scenarios.

In e-commerce, customer reviews wield significant influence over business strategies. This study proposes a robust sentiment analysis (SA) model tailored to e-commerce recommendations. It aims to address the key limitations of existing methods, including challenges in generalizability, feature extraction, class imbalance, and hyperparameter tuning. Our process uses an autoencoder (AE) to extract key features from the input sentence. We employ DistilBERT for word embedding, which performs faster than the standard BERT model (bidirectional encoder representations from transformers). The proposed architecture integrates an AE with a transductive long short-term memory (TLSTM) unit, which is trained with a modified generative adversarial network (GAN). TLSTM leverages transductive learning to emphasize training samples that closely resemble those in the test distribution, enhancing the flexibility and predictive accuracy of the model. Within the GAN, the generator is designed to exclude gradients from dominant batch instances, encouraging greater output diversity and generalization. Once the AE is trained, its compressed feature representations are fed into a multilayer perceptron (MLP) classifier. To tackle class imbalance issues during classification, we implement a reinforcement learning (RL) mechanism. This strategy prioritizes the minority class by applying a reward mechanism to balance the classification outcomes. Moreover, we use the Bayesian optimization hyperband (BOHB) algorithm to fine-tune the hyperparameters of the model. Experimental results on the AIV, AA, and Yelp datasets demonstrate superior performance, with F-measure scores of 91.603%, 89.504%, and 90.397%, respectively. These outcomes validate the robustness of the model and its potential to significantly enhance recommendation quality in dynamic e-commerce environments.