Applied Soft Computing最新文献

{"title":"A universal framework for multi-class imbalanced learning based on Huffman Tree","authors":"Zhong-Liang Zhang,&nbsp;Xiao-Xi Zhao,&nbsp;Heng-Ming Wu,&nbsp;Lei Wang,&nbsp;Xing-Gang Luo","doi":"10.1016/j.asoc.2026.114631","DOIUrl":"10.1016/j.asoc.2026.114631","url":null,"abstract":"<div><div>Compared to binary imbalanced classification tasks, multi-class imbalanced problems face greater challenges due to the presence of multiple classes and significant sample size differences between classes. Furthermore, a unified, generic framework for simplifying multi-class imbalanced problems does not currently exist. In addressing these issues, this paper proposes a universal framework based on the Huffman tree, which effectively transforms complex multi-class imbalanced problems into multiple relatively balanced binary classification sub-problems. It not only simplifies model construction and problem-solving but also reduces class imbalance ratios. Within this framework, a cost-sensitive method is embedded to further address class imbalance within the sub-problems, namely the Huffman Tree-Biased Support Vector Machine (HT-BSVM). The Arithmetic Optimization Algorithm (AOA), a metaheuristic algorithm, is employed to optimize HT-BSVM. Extensive experimental results and statistical tests demonstrate that HT-BSVM significantly reduces the imbalance ratio and exhibits outstanding performance.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114631"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039866","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"One-class anomaly detection based on image reconstruction by Wavelet, Gaussian Fourier and variational Embedding for medical images","authors":"Lin-Chieh Huang,&nbsp;Hung-Hsu Tsai,&nbsp;Yu-Che Chuang","doi":"10.1016/j.asoc.2026.114633","DOIUrl":"10.1016/j.asoc.2026.114633","url":null,"abstract":"<div><div>This paper proposes a new pixel-reconstruction-based method that combines high-frequency sub-bands of <strong>W</strong>avelets, <strong>G</strong>aussian <strong>F</strong>ourier features (GFF) and <strong>V</strong>ariational <strong>E</strong>mbedding (VE) for one-class anomaly detection on medical images, which is called WGF-VAE hereafter. Traditional reconstruction-based methods reconstruct low-frequency information, leading to miss image details during image reconstruction, especially for medical image reconstruction. As a result, those methods often cause false-positives, regarding normal parts as anomalies. The WGF-VAE scheme can overcome these drawbacks mentioned above due to the use of high-frequency sub-bands of wavelets, GFF and VE in the design of image reconstruction process. High-frequency sub-bands conserve high-frequency information, making the decoder of the WGF-VAE scheme easier to learn and handle these details of images. Moreover, the decoder leverages GFF to cover a broader frequency spectrum by transforming coordinates of an input image into a higher-dimension space so as to enhance the learning of high-frequency functions. Meanwhile, the scheme can accurately capture and reconstruct high-frequency details of medical images by utilizing the localized frequency information from high-frequency sub-bands and the expanded frequency spectrum from GFF. Furthermore, a variational autoencoder (VAE) produces VE which is employed in the decoding phase to play a role as the latent feature of high-frequency sub-bands. It makes the decoder stable to yield normal images so as to precisely compute the difference between input and output images, resulting in promoting the recognition ability of the scheme. Hence, the WGF-VAE scheme possesses remarkably ability on detection and localization for anomalies because of taking a combination of three features as inputs of the decoder. Finally, massively experimental results show that the WGF-VAE scheme outstandingly surpasses state-of-the-art methods on anomaly detection for brain and liver images in two public benchmarks.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114633"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039914","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-level retrieval with representation alignment enhances cross-document evidence synthesis for scientific knowledge generation","authors":"Shenghua Wang ,&nbsp;Zhongqiang Wang ,&nbsp;Chunhe Qu ,&nbsp;Zhen Yin","doi":"10.1016/j.asoc.2026.114649","DOIUrl":"10.1016/j.asoc.2026.114649","url":null,"abstract":"<div><div>Recent advancements in retrieval-augmented generation (RAG) have significantly improved performance in knowledge-intensive natural language processing tasks. However, applying RAG to scientific literature for knowledge synthesis remains challenging due to the difficulty in effectively retrieving and integrating multi-element evidence across documents and in modeling long-range dependencies. Moreover, the absence of dedicated evaluation resources for multi-element, cross-document scientific RAG further limits progress in this area. To address these challenges, this paper proposes a novel approach that integrates multi-level retrieval and contrastive learning to enhance RAG for scientific literature. Specifically, contrastive learning is employed to optimize embeddings for retrieving diverse element types, including text, tables, formulas, and figures, while a multi-level retrieval strategy combines coarse-grained document filtering with fine-grained element ranking to improve retrieval accuracy and efficiency. In addition, Chain-of-Thought (CoT) reasoning is incorporated to enhance the coherence of synthesized knowledge. To support research in this direction, we construct SciLitQA, a new dataset comprising over 50,000 high-quality Q&amp;A pairs tailored for evaluating multi-element, cross-document knowledge synthesis. Extensive experiments on SciLitQA and public benchmarks demonstrate the effectiveness of our approach and establish SciLitQA as a valuable resource for advancing research in scientific RAG.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114649"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039912","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Self-supervised inference of missing energy sources for enhancing energy consumption forecasting","authors":"Jong Seong Park,&nbsp;Jeong-Ha Park,&nbsp;Ji-Hyeok Choi,&nbsp;Hyuk-Yoon Kwon","doi":"10.1016/j.asoc.2026.114732","DOIUrl":"10.1016/j.asoc.2026.114732","url":null,"abstract":"<div><div>While incorporating multiple correlated energy sources significantly improves energy consumption forecasting accuracy, practical data collection limitations often prevent access to complete multi-energy datasets across all regions. This paper addresses the critical gap between the theoretical benefits of multi-energy forecasting and the practical constraints of incomplete data collection by introducing a novel cross-regional inference framework. We propose the first self-supervised learning approach for inferring missing energy sources in multi-energy forecasting, which leverages complete datasets from reference regions to infer missing energy consumption patterns in target regions with incomplete data collection. The key innovation lies in our self-supervised representation learning method that captures complex inter-energy correlations beyond traditional distance-based metrics, enabling effective clustering of households with similar multi-energy consumption patterns. Our approach provides theoretical guarantees for clustering convergence and includes comprehensive complexity analysis, addressing fundamental questions about the scalability and reliability of cross-regional energy inference. Through extensive experiments on real-world datasets, we demonstrate substantial improvements over baseline approaches. Specifically, our method achieves up to 30% better inference accuracy compared to simple averaging methods. Furthermore, in the downstream forecasting task, it reduces the forecasting error (MAE) by an average of 15.2% and a maximum of 27.0% compared to single-source baselines, showing robust performance across different temporal patterns and regional characteristics. The source code for the proposed framework is available at <span><span>Anonymous GitHub Repo</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114732"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190708","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MAGF-CCL: Multi-level attentive graph fusion with cross-modal complementary learning for internal control material weaknesses prediction","authors":"Xuan Zhang ,&nbsp;Boyu Hu ,&nbsp;Xusheng Sun ,&nbsp;Jingling Ma ,&nbsp;Gang Wang ,&nbsp;Tingting (Rachel) Chung","doi":"10.1016/j.asoc.2026.114656","DOIUrl":"10.1016/j.asoc.2026.114656","url":null,"abstract":"<div><div>Internal control material weaknesses (ICMW) are often early warnings of possible financial misstatements or fraud that can lead to financial distress. Therefore, accurately predicting ICMW is crucial to mitigating greater losses. Recent studies have shown that multi-modal data holds significant promise for predicting ICMW in listed companies. However, the complementary effects of multi-modal data remain underexplored. This limits the model’s ability to fully capture the ICMW clues. Furthermore, existing studies primarily treat companies as independent entities. They overlooked the inter-company relationships that may influence the final prediction results. To address these limitations, this study proposes a Multi-level Attentive Graph Fusion with Cross-modal Complementary Learning (MAGF-CCL) method for ICMW prediction. Specifically, first, the instance-level graphs are constructed using k-nearest neighbors (KNN) algorithm. Graph Convolutional Network (GCN) is then employed to learn inter-company relationships in graphs. Second, a Multi-modal Complementary Learning (MCL) module is designed to explore the multi-modal complementarity, hence fully capturing ICMW clues. Third, to integrate multi-modalities effectively, the numerical and textual graphs are fused using Modality-level Fusion Mechanism (MFM) and Structure-level Fusion Mechanism (SFM). These fusion modules combine the multi-modal data and structural relationships, respectively. Finally, the fused graph is subsequently fed into a GCN to facilitate cross-modal information propagation and enhance ICMW prediction. Experimental results on a real-world dataset demonstrate that the proposed MAGF-CCL method outperforms state-of-the-art (SOTA) methods in predicting ICMW. The AUC value of MAGF-CCL achieved 91.04 %, outperforming existing SOTA method by nearly 3 %. This study also visualized the inter-company relationships and attention maps of MFM module, thereby providing relevant decision support for stakeholders.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114656"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145986733","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"From reading minds to books: A dual-sided multimodal knowledge-aware recommender system based on collaborative contrastive learning","authors":"Zhenyu Li ,&nbsp;Shiwei Wang ,&nbsp;Zhaodong Ding ,&nbsp;Chanyuan Zhang ,&nbsp;Xin Yan ,&nbsp;Zongfeng Zou","doi":"10.1016/j.asoc.2026.114744","DOIUrl":"10.1016/j.asoc.2026.114744","url":null,"abstract":"<div><div>Existing studies on multimodal recommender systems primarily focus on the fusion of multimodal features from items, largely overlooking the exploration of users’ multimodal information. Additionally, the effectiveness of recommender systems is often limited in real-world applications due to cold-start and long-tail problems. To address these challenges, we propose a dual-sided multimodal recommender system named UAFGK (<strong>U</strong>ser-<strong>A</strong>ware <strong>F</strong>ine-<strong>G</strong>rained <strong>K</strong>nowledge) in the context of book recommendation. Firstly, we construct multi-view graphs incorporating multimodal information from both the user and item sides. Specifically, we design a fine-grained knowledge selector using pairwise TF-IDF and pre-trained models to build a fine-grained knowledge hypergraph of items. Meanwhile, users’ modalities are utilized to construct visual and textual similarity graphs of users. Secondly, to consider the different high-order relationships in multi-view graphs, UAFGK models these relationships through hypergraph convolution networks (HGCNs) and graph convolutional networks (GCNs) in two distinct ways. Finally, to enhance the performance of the recommender system, we introduce a collaborative contrastive learning (CL) method that incorporates tight CL, loose CL, BPR, and cross-view BPR. Extensive experimental results on two novel datasets (DoubanBook-60K and DoubanMovie-200K) curated from community platforms demonstrate that UAFGK consistently and significantly outperforms state-of-the-art (SOTA) models across all metrics. Additionally, each component is validated as effective through detailed ablation studies. By focusing on the content similarities of users and items, our approach effectively mitigates cold-start and long-tail challenges to a certain extent. UAFGK can capture knowledge-aware user preferences and offer explainability to multimodal recommender systems.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114744"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191515","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A MIP-informed machine learning method for the real time management of a multi-storage microgrid with renewable energy sources","authors":"Stanislav Fedorov,&nbsp;Anna Pinnarelli,&nbsp;Pasquale Vizza,&nbsp;Maria Elena Bruni","doi":"10.1016/j.asoc.2026.114720","DOIUrl":"10.1016/j.asoc.2026.114720","url":null,"abstract":"<div><div>Complex microgrids with the inclusion of renewable energy sources and multiple energy storage systems based on batteries and hydrogen storage require fast, reliable, and adaptable management approaches to achieve profitability. Uncertainty in power generation and demand levels renders the application of typical optimization methodologies impossible in real-time management scenarios, due to slow response times and the risk of power imbalances. In this paper, we introduce a new machine learning based approach to the real-time management of complex microgrids under uncertainty. A decision tree classifier is trained on solved-to-optimality mixed-integer linear programming problem instances. Then, a decision algorithm is proposed to dispatch the microgrid operation. The proposed machine learning approach shows promising performance on a real dataset simulation compared to the benchmark rule-based approach, resulting in a significant 22% yearly operational cost reduction tested on the real data.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114720"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191517","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Histopathological breast cancer image classification using adaptive hunt optimization-based deep convolutional neural network","authors":"Afnan M. Alhassan","doi":"10.1016/j.asoc.2026.114729","DOIUrl":"10.1016/j.asoc.2026.114729","url":null,"abstract":"<div><div>Breast Cancer is a serious disease and one of the most common cancers among women globally. However, survival rates are high with early diagnosis and comprehensive treatment. Though the diagnosis of Breast Cancer (BC) at an early stage boosts survivability, the classification of the tumor or cancer cells is intricate. The classification of tumors based on Histopathological images (HPIs) as benign or malignant paved the way for more effective and spontaneous treatment. Even though researchers developed numerous techniques for BC classification, they still struggle with limited resources, privacy preservation, high computation cost, overfitting issues, poor generalizability, and interpretability. To overcome these limitations, an Adaptive Hunt Optimization-based Deep Convolutional Neural Network (AHO-based Deep CNN) model is presented. The model incorporates the Deep CNN with the Adaptive Hunt optimization (AHO), which enhances the convergence rate while minimizing loss. The inclusion of a federated Generative Adversarial Network (GAN) based Segmentation process identifies and segments the tumor region, thereby facilitating more accurate BC classification while considering privacy concerns. More specifically, the proposed AHO-based Deep CNN is evaluated against other existing models based on multiple datasets to enhance the performance. Specifically, the proposed model demonstrated high performance on the BreakHis dataset, achieving an Accuracy of 94.25 %, a kappa of 0.91, a precision of 94.48 %, a Sensitivity of 96.71 %, and a Specificity of 83.06 in BC classification with 90 % of the data used for training. Overall, the central objective of this study is to develop a highly effective deep learning model for multiclass BC classification to improve the prognosis and survival rate of patients.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114729"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191521","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A consensus particle swarm–guided trajectory unification method for training Kolmogorov–Arnold Networks","authors":"Zhihao Liu ,&nbsp;Yongfeng Zhang","doi":"10.1016/j.asoc.2026.114769","DOIUrl":"10.1016/j.asoc.2026.114769","url":null,"abstract":"<div><div>We propose CPSOTJU, a novel two-stage hybrid optimization framework designed to address slow convergence and local optimum entrapment in Kolmogorov–Arnold Networks (KANs). CPSOTJU integrates the global exploration capability of Consensus Particle Swarm Optimization (CPSO) with the precise local refinement of Trajectory Unification (TJU). In the first stage, CPSO explores the high-dimensional spline parameter space to identify promising basins of attraction, thereby mitigating premature convergence. In the second stage, TJU performs trajectory-guided local refinement to obtain stable, high-quality solutions. Extensive experiments on eight analytic benchmark functions and the Energy Efficiency regression dataset demonstrate that CPSOTJU consistently outperforms gradient-based baselines and derivative-free competitors such as CMA-ES, while remaining competitive with strong adaptive optimizers like AdamW. In particular, CPSOTJU achieves relative test RMSE reductions of up to 84.0% compared to CMA-ES and an average reduction of 66.02% across complex multimodal benchmarks relative to first-order methods, confirming its robustness and efficiency for training KANs.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114769"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146190765","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Saturation and pressure change estimation from time-lapse seismic data using vision transformers","authors":"Marcos Cirne ,&nbsp;Masoud Maleki ,&nbsp;Michael Diniz ,&nbsp;Denis Schiozer ,&nbsp;Alessandra Davolio ,&nbsp;Anderson Rocha","doi":"10.1016/j.asoc.2026.114715","DOIUrl":"10.1016/j.asoc.2026.114715","url":null,"abstract":"<div><div>Deep learning-based algorithms have been used in oil and gas industry applications to optimize the performance of four-dimensional (4D) seismic inversion methods. These pose an important problem in reservoir management since they convey substantial information about fluid recovery and distribution on the reservoir over time. In this work, we propose a novel deep learning architecture based in Vision Transformers (ViT) to estimate saturation and pressure changes from 4D seismic data using data collected from a pre-salt oil field. The proposed method is first trained with an ensemble of simulation models (synthetic domain) and the resulting architecture is used to predict the actual changes using real 4D seismic data (observed domain). Due to the absence of data points around the reservoir area conveyed in our dataset, we redesigned the ViT architecture so that it only processes the existing data points in a 1-D fashion, making the architecture capable of processing different reservoir shapes other than rectangular ones. The quantitative results against recent deep learning methods corroborate the effectiveness of our proposed architecture. Finally, a qualitative analysis by a specialist and the use of an Explainable Artificial Intelligence (XAI) method helped assess the key differences in the decision-making processes taken by each analyzed algorithm.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114715"},"PeriodicalIF":6.6,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080101","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}