FTA2C: Achieving superior trade-off between accuracy and robustness in adversarial training.

HEFT-DETR: A chip defect detection method based on hierarchical edge-Sensing feature transformation technology

Glandular segmentation holds significant importance in the pathological analysis of hematoxylin and eosin (H&E) stained images, particularly in the diagnosis and treatment of breast and colorectal cancers. Accurate segmentation of glandular structures provides critical support for lesion detection and pathological assessment. Traditional fully supervised learning methods typically require a substantial amount of labeled data, which is often challenging to obtain in the medical field. To address this issue, this study proposes a novel weakly supervised segmentation method that integrates pseudo-labels and consistency contrast loss with random noise, achieving results comparable to several state-of-the-art architectures trained under full supervision. The proposed network model, LGDUnet, features both global and local branch structures. The global branch focuses on the overall structure of the glands, while the local branch emphasizes the capture of fine edge features. The encoder utilizes a combination of ResHorBlock and WSAB to enhance feature extraction capabilities through higher-order interactions and weighted spatial attention at various levels. The decoder employs the Dual Self-Attention Transpose (DST) structure, which enhances reconstruction accuracy through a dual self-attention mechanism. Skip connections are implemented using the Dual Attention Transformer (DAT) for encoder feature transformation, further improving the efficacy of feature propagation.We conducted comprehensive comparative and ablation experiments on the benchmark dataset GlaS from the MICCAI 2015 Challenge, our self-constructed breast cancer dataset Tubule of Breast Cancer (TBC), and the Colorectal Cancer Gland Dataset (CGD). The experimental results demonstrate that LGDUnet exhibits superior performance in glandular segmentation tasks, validating the effectiveness of this method within the weakly supervised learning framework.

Wearable Human Activity Recognition (HAR) models often degrade across users and sensor placements due to domain shifts. This paper presents the Multi-Sensor Adaptive Feature Alignment Network (MSAFAN), integrating Sensor-Specific Normalization Layer (SSNL), Hybrid Polynomial Feature Transformation (HPFT), Conditional Alignment Loss (CAL), and Entropy-Guided Pseudo-Labeling (EGPL) for class-wise adaptation and robust cross-sensor generalization. Evaluated on four benchmark datasets: BAR, DSADS, PAMAP2, and MM-DOS, the MSAFAN improves macro-F1 by 8.4% and accuracy by 10.3% while reducing computational cost by 26% over state-of-the-art UDA models. The framework achieves stable convergence, efficient adaptation, and scalable performance, confirming its suitability for real-time deployment in edge AI and wearable computing applications.

Hyperspectral images present significant challenges for conventional dimensionality reduction methods due to their high dimensionality, spectral redundancy, and complex spatial–spatial dependencies. While kernel-based sparse representation methods have shown promise in handling spectral non-linearities, they often fail to preserve spatial consistency and semantic discriminability during feature transformation. To address these limitations, we propose a novel semantic-guided kernel low-rank sparse preserving projection (SKLSPP) framework. Unlike previous approaches that primarily focus on spectral information, our method introduces three key innovations: a semantic-aware kernel representation that maintains discriminability through label constraints, a spatially adaptive manifold regularization term that preserves local pixel affinities in the reduced subspace, and an efficient optimization framework that jointly learns sparse codes and projection matrices. Extensive experiments on benchmark datasets demonstrate that SKLSPP achieves superior performance compared to state-of-the-art methods, showing enhanced feature discrimination, reduced redundancy, and improved robustness to noise while maintaining spatial coherence in the dimensionality-reduced features.

Mamba-enhanced disease semantic knowledge graph for interpretable automatic ICD coding.

Forecasting financial time series is particularly challenging since market data is complicated and non-stationary, and it is necessary to identify both short-term momentum and long-term structural patterns. This work develops the Dynamic Group Fusion Enhanced Transformer (DGFET), a new approach that integrates adaptive feature group fusion and selective information processing. The suggested DGFET architecture has Group-FiLM adapters that use Dynamic Group Fusion techniques for adaptive feature transformation to manage market, fundamental, technical, and sentiment feature groups. We assess the model using four unique labeling strategies: short-horizon momentum (binary/ternary) and triple-barrier (binary/ternary), which represent various temporal horizons and forecasting methods. Our ablation analysis, conducted on a comprehensive EUR/USD dataset from 2010 to 2023 with 88 features, demonstrates that the proposed method consistently outperforms baseline LSTM and standard transformer models across all prediction objectives. The improved architecture has a higher overall performance, with an F1-macro score of 0.5356 and a ROC-AUC of 0.6612. It also works very well for short-horizon momentum binary classification (F1: 0.7219, ROC-AUC: 0.8105). The results show that adaptive feature fusion works better than traditional designs when combined with dynamic group selection. The best configurations depend on the specific prediction job. Our results underscore the imperative of task-specific architectural design in financial machine learning applications, especially for methodologies necessitating varied temporal horizons and prediction granularities.

Abstract In the domain of multi-modal perception, integrating visual and tactile information enhances robot interaction and scene comprehension. Existing fusion methods lack cross-modal alignment mechanisms in intermediate layers. They rely on fixed strategies, neglecting inherent modality correlations and preventing noise suppression. To address these limitations, we propose a hierarchical multi-scale architecture. This framework enables multiscale feature fusion through inter-level connections, generating modality-consistent representations. These representations serve as query vectors in focal modulation. Parameterized modulators dynamically calibrate single-modal embeddings to enhance semantic alignment. A dual-path attention mechanism with dynamic gating enables adaptive feature transformation. It selectively integrates modalities based on reliability. Experimental results show that FM-Net consistently outperforms existing methods on ObjectFolder 2.0, AU, and FMQ datasets. Code is available at: https://github.com/lifuqin258/FM-Net.git.Convolutions NetworkRecent advances in convolutional networks have significantly improved feature representation capabilities. Hu et al.[9] developed the Squeeze-and-Excitation Network (SENet), which uses channel attention to adaptively recalibrate feature importance. Li et al.[10] introduced Selective Kernel Network (SKNet), employing multi-scale feature selection to capture spatial details at varying resolutions. Woo et al.[11] proposed the Convolutional Block Attention Module (CBAM), integrating channel and spatial attention mechanisms for comprehensive feature refinement. Hou et al.[12] presented Coordinate Attention (CA), which incorporates positional information to enhance long-range dependencies while maintaining computational efficiency. Ouyang et al.[13] designed the Efficient Multi-Scale Attention (EMA) module, optimizing channel information retention through a parameter-efficient multi-scale mechanism.Prevailing convolutional network architectures, while demonstrating strong performance for unimodal feature learning, exhibit inherent limitations in cross-modal assimilation. Our approach addresses this limitation through a multi-resolution convolutional framework that establishes

Abstract: Capital market or financial market volatility is pervasive and complex. It is the pulse of the market that reflects balance between opportunity and uncertainty. In the fast-growing Indian stock market, accurately modelling and predicting this volatility is vital for achieving superior performance. Traditional econometric models like ARCH and GARCH have offered fundamental insights but fail to explain the nonlinear and regime shifting patterns of volatility. This paper introduces AI Volatility Adaption Cycle (AIVAC) i.e., a conceptual framework that integrates volatility theory, complexity economics and artificial intelligence to build adaptive trading strategies. The AIVAC blends econometric reasoning with neural and reinforcement learning to form a self-evolving intelligent framework. The proposed framework has five layers namely, data fusion and conceptualization, feature transformation, predictive intelligence and regime detection, adaptive strategy engine and feedback driven evolutionary learning. This will allow the users to automatically adjust their settings like risk levels and trading methods on real time basis. This paper contributes to framing a practically applicable adaptive intelligent framework fit for Indian market. The paper also discusses the implications for the stakeholders while emphasizing transparency, interpretability and ethical AI adoption. Keywords: Volatility, Deep Learning, Adaptive Trading, Artificial Intelligence, Indian Stock Market

Accurate prediction of biofuel yield and combustion emissions plays a key role in improving conversion efficiency and reducing dependence on trial-and-error experiments. Biofuel systems involve diverse biomass feedstocks and complex thermochemical and combustion processes, which makes modeling difficult. Reliable prediction tools also support cleaner energy practices and informed process control. Existing research shows several clear limitations. Many studies rely on small, single-site datasets, which limit broader applicability. Data preprocessing methods differ widely across publications, leading to inconsistencies in reported results. Validation strategies are often limited to internal testing, which restricts confidence in real-world use. These issues reduce model generalization, reproducibility, and clarity of interpretation. This review examines recent progress in artificial intelligence and machine learning applied to biofuel production and engine emission prediction. It summarizes commonly used data sources, including laboratory experiments and engine tests. The review outlines feature selection and transformation methods adopted in prior work. It also reviews model construction strategies and evaluation practices used to assess performance. Surveyed studies show that ensemble learning methods, neural networks, and physics-informed hybrid models achieve high prediction accuracy at laboratory scale. These models perform well for yield and emission estimation under controlled conditions. At the same time, several persistent challenges remain. Many advanced models show weak extrapolation beyond training ranges. Model transparency is also limited, which affects trust and interpretability. The findings indicate that benchmark datasets and consistent preprocessing protocols are needed. Strong external validation can improve reliability. Incorporating physical constraints into machine learning workflows can enhance stability and realism. Such practices can support real-time implementation and promote wider use of data-driven prediction tools in biofuel research and industrial operations.

Abstract The progressive growth of seismological databases has motivated the exploration of novel methodologies for common tasks such as detection and phase-picking, with a focus on maintaining reliability comparable to human performance. This goal consistently involves leveraging deep learning techniques, which emulate sensory processing in the human brain through numerical simulations. This study introduces a physically driven feature engineering approach that capitalizes on the inherent information within seismic data. While many contemporary studies train their models via robust raw datasets, practical alternatives tailored for smaller databases are often overlooked. Feature engineering in seismological contexts aims to develop deep learning models with tangible physical significance, specifically those that target event detection and phase-picking tasks across both local and regional seismic environments. Our approach leverages physically driven feature transformations for the joint detection and phase-picking task. This includes incorporating the energy signal envelope for effective seismic event classification, using amplitude spectra from signals filtered at predefined frequency bands, and calculating spatial features (such as wave incidence and azimuth) for accurate phase-picking. This integrated feature set optimizes model performance, especially when dealing with small volcanic seismology datasets. The proposed joint methodology is particularly pertinent in seismo-volcanic contexts, where accurate discrimination and characterization of seismic signals are pivotal for monitoring and risk assessment purposes. The incorporation of significant physical information from seismic signals into pattern recognition is crucial, as many feature engineering applications lack a contextual understanding of the data, which can lead to distortions, particularly within geophysical domains. Our results demonstrate human-level performance in these common tasks, harnessing the capabilities of statistical learning algorithms as a practical, resource-efficient solution for addressing these challenges on a large scale.

Many deep learning-based spatiotemporal fusion (STF) methods have been proven to achieve high accuracy and robustness. Due to the variable shapes and sizes of objects in remote sensing images, pyramid networks are generally introduced to extract multi-scale features. However, the down-sampling operation in the pyramid structure may lead to the loss of image detail information, affecting the model’s ability to reconstruct fine-grained targets. To address this issue, we propose a novel Dual-Stream Enhanced Pyramid based on Generative Adversarial Network (DSEPGAN) for the spatiotemporal fusion of remote sensing images. The network adopts a dual-stream architecture to separately process coarse and fine images, tailoring feature extraction to their respective characteristics: coarse images provide temporal dynamics, while fine images contain rich spatial details. A reversible feature transformation is embedded in the pyramid feature extraction stage to preserve high-frequency information, and a fusion module employing large-kernel and depthwise separable convolutions captures long-range dependencies across inputs. To further enhance realism and detail fidelity, adversarial training encourages the network to generate sharper and more visually convincing fusion results. The proposed DSEPGAN is compared with widely used and state-of-the-art STF models in three publicly available datasets. The results illustrate that DSEPGAN achieves superior performance across various evaluation metrics, highlighting its notable advantages for predicting seasonal variations in highly heterogeneous regions and abrupt changes in land use.

Implementation of new normalisation-based feature transformation methodology for multi-task learning using residual TCN with multi-channel and spatial attention

The growth of the automotive industry in Indonesia has contributed to high demand for used cars as a more economical alternative to new cars. However, determining the price of used cars is often a challenge for showrooms and prospective buyers because it involves many factors and is subjective. This study aims to develop a used car price prediction model using the Support Vector Regression (SVR) algorithm with a Radial Basis Function (RBF) kernel approach. A total of 1,000 entries were obtained through web scraping from the cintamobil.com website. The research methodology refers to the CRISP-DM framework, starting from business understanding to model deployment through a web application using Streamlit. The preprocessing process involves handling missing values, outliers, data duplication, and numerical and categorical feature transformations. The SVR model was evaluated using RMSE, MAPE, and MAE metrics to assess prediction accuracy. The results show that SVR is capable of providing fairly accurate price predictions, with parameters C=1, gamma=0.1, and epsilon=0.1 producing the best performance, namely an MAE value of IDR 6,472,572, an RMSE of IDR 8,958,555, and a MAPE of 3.41%. Referring to the prediction accuracy category based on the MAPE value, where a MAPE value ≤ 10% is categorized as high accuracy, it can be concluded that this model has high prediction accuracy. This shows that the SVR model used is capable of estimating used car prices with a low error rate and good accuracy.

Abstract Introduction Precision oncology, which matches patients with optimal treatments, currently benefits only a minority of patients. This limited impact stems from two challenges: (i) only a small set of genomic alterations are actionable, and (ii) oncogenic mutations do not reliably predict tumor dependency. In heterogeneous tumors, drugs may eliminate only a subset of cells, while resistant populations persist and drive recurrence. To address these gaps, we present DeepVul, a multi-task transformer that jointly predicts gene essentiality and drug response from transcriptomes. By aligning gene expression with genetic and pharmacologic perturbations in a shared latent space, DeepVul predicts cancer-cell vulnerabilities to many genes and drugs from a single molecular readout, rendering whole-genome transcriptomes clinically actionable. Benchmarking showed that DeepVul matches or complements mutation-based approaches and, through interpretability, identifies mechanisms of response and resistance. DeepVul is publicly available at https://github.com/alaaj27/DeepVul.git. Methods DeepVul learns a latent representation linking gene expression to essentiality and perturbation data via a multi-task objective; a second objective fine-tunes drug response by minimizing per-drug MSE. The architecture comprises feature transformation, encoder-based feature extraction, and optional fine-tuning (frozen vs. tunable encoder). We evaluate on held-out cohorts with ablations against baselines. Results DeepVul matched or outperformed baselines in per-gene correlation and correctly predicted essentiality for 46% of genes with correlation >0.3, with the largest gains on actionable oncogenes. Trained on DepMap and applied without retraining, it generalized to an independent Sanger cohort, whereas the baseline model failed to generalize. For drug response, transfer from essentiality with a frozen encoder outperformed fine-tuning. Compared with mutation-based rules, DeepVul identified populations sensitive to targeted therapies that showed significantly lower experimental essentiality across 32 actionable genes (p<0.05); the BRAF-only analysis showed the same trend but was not significant due to limited sample size. SHAP/LISA recovered known resistance biology, including STAT3-mediated resistance to BRAF inhibitors. Conclusions DeepVul predicts gene essentiality and drug response from expression profiles, matches or exceeds baselines, and generalizes across cohorts without retraining—an important requirement for clinical translation. Using BRAF and 32 actionable genes as use cases, it identifies treatment-sensitive cell lines with accuracy comparable to mutation-based precision oncology while extending coverage beyond the few currently actionable genes. DeepVul complements mutation-based strategies by providing scalable, transcriptome-driven predictions of therapeutic vulnerabilities and may inform transcriptome-based clinical decision support. Citation Format: My Bach Nguyen, Ala Jararweh, David Arredondo, Oladimeji Macaulay, Luis Tafoya, Yue Hu, Avinash Sahu, Genevieve Boland, Keith Flaherty, Mikaela Dicome. DeepVul: A multi-task transformer model for joint prediction of gene essentiality and drug response [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A024 .

Classification of bacterial species, particularly for closely related taxa, remains a major challenge in many areas, e.g., public health, food industries, and many others. The issues are mainly caused by overlapping genetic features of organisms and data complexities. In this study, a bacterial taxonomic identification framework that integrates genome-derived motif sequences with machine learning was introduced. Two hundred and forty genome sequences from Salmonella enterica, representing six subspecies and ten serovars, were used for modelling. Sequence motifs were predicted from single-copy orthologous core genes of the downloaded genomes. Single nucleotide polymorphisms (SNPs) within these motifs were extracted and numerically encoded as machine learning features. The 20 top-most informative predictors from feature selections were used for model training in Random Forest and Support Vector Machine. Comparing the output from multiple analyses, the Random Forest model achieved the highest accuracy of 97.92%, demonstrating reliable differentiation of Salmonella at both subspecies and serovar levels. This research presents two key innovations: i) the use of sequence motifs as molecular signatures for bacterial classification; ii) a novel feature engineering method that transforms genome-derived data into machine learning-readable features. The proposed framework offers a practical and scalable solution for fine-level bacterial classification and has high potential to be applied for other microbial taxa.

This study aims to develop an optimized machine learning-based sentiment classification model for election-related issues. A dataset comprising 10,001 entries was collected from the social media platform X and manually labeled into three sentiment classes: positive, negative, and neutral. The preprocessing stage involved text cleaning, stemming, and feature transformation using the Term Frequency-Inverse Document Frequency (TF-IDF) method. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was employed. Three baseline classification algorithms—K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Gaussian Naive Bayes (GNB)—were initially evaluated to establish a performance benchmark. Model development proceeded by applying hyperparameter optimization using the Optuna framework and further enhancing the models via boosting with Extreme Gradient Boosting (XGBoost). Experimental results revealed that the combination of SVM with Optuna and XGBoost achieved the best performance, reaching 97% accuracy, precision, recall, and F1-score across all classes. In contrast, the KNN and GNB models experienced a notable decline in performance following hyperparameter tuning, although partial recovery was observed when combined with boosting. These findings suggest that hyperparameter tuning and boosting are not universally effective across all classifiers, yet their synergistic application significantly enhances performance in SVM-based models. This study highlights the importance of model-specific optimization strategies in building robust sentiment analysis systems, particularly for handling unbalanced public opinion data in social media contexts.

ABSTRACT Objective Oral potentially malignant disorders (OPMDs) are defined as clinical conditions that carry the risk of cancer development in the oral cavity. Regardless of their cancerous nature, the mechanism of malignant transformation is unknown and there are no indicators to predict the risk of malignant transformation. In order to manage OPMDs appropriately, indicators that are useful for predicting their prognosis are required. The purpose of this study was to examine the expression of YAP1, p53, and Ki‐67 in OPMDs and explore their potential as biomarkers to predict the risk of malignant transformation. Methods Immunohistochemical staining for YAP1, p53, and Ki‐67 was performed in 70 cases with a confirmed diagnosis of OPMDs. The staining status was evaluated by multiplying the positive cell occupancy score by the staining intensity score. We statistically investigated the association between malignant transformation of OPMDs and clinicopathologic features and staining grade. Results No significant differences were observed between malignant transformation and clinicopathological features of OPMDs. However, YAP1 ( p = 0.003) and p53 ( p = 0.013) were expressed higher in malignant transformation cases than in non‐malignant transformation cases. Moreover, Ki‐67 showed a tendency to be highly expressed in malignant transformation cases ( p = 0.161). Conclusions The cancer‐related factors YAP1 and p53 were found to be potentially upregulated in cases of malignant transformation of OPMDs. Therefore, YAP1 and p53 may be biomarkers to predict malignant transformation of OPMDs, and patients with OPMDs with high expression of YAP1 and p53 are at high risk of malignant transformation and require careful follow‐up.

Many network architectures exist for learning on meshes, yet their constructions entail delicate trade-offs between difficulty learning high-frequency features, insufficient receptive field, sensitivity to discretization, and inefficient computational overhead. Drawing from classic local-global approaches in mesh processing, we introduce PoissonNet, a novel neural architecture that overcomes all of these deficiencies by formulating a local-global learning scheme, which uses Poisson's equation as the primary mechanism for feature propagation. Our core network block is simple; we apply learned local feature transformations in the gradient domain of the mesh, then solve a Poisson system to propagate scalar feature updates across the surface globally. Our local-global learning framework preserves the features's full frequency spectrum and provides a truly global receptive field, while remaining agnostic to mesh triangulation. Our construction is efficient, requiring far less compute overhead than comparable methods, which enables scalability—both in the size of our datasets, and the size of individual training samples. These qualities are validated on various experiments where, compared to previous intrinsic architectures, we attain state-of-the-art performance on semantic segmentation and parameterizing highly-detailed animated surfaces. Finally, as a central application of PoissonNet, we show its ability to learn deformations, significantly outperforming state-of-the-art architectures that learn on surfaces. https://github.com/ArmanMaesumi/poissonnet

Accurate traffic accident risk prediction is crucial for enhancing urban road network efficiency and safety, attracting increasing attention in traffic forecasting research. However, existing models often struggle to capture the global spatial-temporal correlations and spatial heterogeneity of traffic accident risk data. They are sensitive to data sparsity, especially for fine-grained prediction tasks. In this paper, we propose a novel spatial-temporal deep neural network, named channel attention-level spatial-temporal convolutional neural network (CA-STNet), to address these challenges. Specifically, to alleviate the impact of data sparsity on the prediction performance of the model, we designed a hierarchical spatial-temporal feature learning framework to capture coarse-grained and fine-grained traffic accident risk characteristics, respectively, and realize cross-scale traffic accident risk feature fusion through a feature transformation matrix, combined with weighted loss function to solve the zero-inflated issue. To better capture the spatial-temporal correlation and channel heterogeneity of traffic accident risk data, a channel-independent self-attention unit is introduced to dynamically capture global spatial-temporal correlation. At the same time, an inter-channel attention unit is adopted to quantify and adjust the importance of different channel characteristics in the spatial dimension. The results of the two real traffic accident data sets indicate that this model outperforms other benchmark models in predictive accuracy. The source code is available at https://github.com/MrYning/CA-STNet .