Traditional regression analysis assumes data homogeneity and employs symmetric loss functions, inherently underperform when confronted with long-tailed or severely imbalanced response distributions. In such scenarios, models are disproportionately dominated by dense regions, leading to significant estimation bias and degraded predictive accuracy in sparse regions where critical samples often reside. Current robust regression approaches frequently exhibit suboptimal performance due to insufficient exploitation of local density structures. To overcome these limitations, we propose a robust BLINEX-SVR framework centered on a response-density-based asymmetric penalty mechanism. Specifically, kernel density estimation is employed to partition samples into distinct response density intervals. The framework employs the Bounded Linear-Exponential (BLINEX) loss function to implement differentiated penalization. Specifically, it assigns exponential penalties to samples situated in sparse response regions to improve long-tail fitting, while applying linear penalties to those in dense response regions to maintain global distributional stability. Unlike conventional Linear-Exponential(LINEX) losses, where unbounded exponential terms often trigger gradient instability and hyperplane oscillations due to Outliers in sparse response regions, BLINEX employs bounding parameters to truncate penalty magnitudes and ensure training stability. This mechanism effectively attenuates the influence of noise and outliers, endowing the model with robust-regression-like stability without compromising its inherent asymmetry. For efficient training, we adapted a Nesterov Accelerated Gradient (NAG) algorithm. Extensive evaluations on ten real-world benchmarks demonstrate that BLINEX-SVR significantly outperforms state-of-the-art methods in both prediction error minimization and overall stability.
Portfolio optimization is an integral domain of study in quantitative finance and gained significant attention from researchers and industry professionals. Deep Reinforcement Learning (DRL) has emerged as a powerful tool for dynamically managing asset allocations aligning with investors’ objectives. However, the critical aspect of asset pre-selection is overlooked. This paper presents a new hybrid approach that combines multi-level return prediction, volatility forecasting, and DRL-based portfolio optimization. We use four predictive models to predict returns: Random Forest (RF), Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). A GARCH model is used to estimate asset volatility. To enhance selection robustness, a novel scoring function is introduced that incorporates predicted returns, momentum, earning growth, and volatility. The weights of these components are adaptively learned using a regime-aware Bayesian optimization strategy. The top-scoring assets are pre-selected at each period and passed to the DRL agent, enabling a dynamic, context-aware portfolio construction. In addition, the portfolio weights are rebalanced at the end of each period using the proposed dynamic rebalancing strategy. The efficacy of our proposed approach is validated using stocks from the NASDAQ-100 and NIFTY-100 indices. The empirical results demonstrated the effectiveness of the proposed methodology based on several performance metrics when compared with benchmarks.
This article introduces generalized two-parameter estimators for modelling explanatory variables in positively skewed data using the inverse Gaussian regression model. The standard approach for estimating unknown regression coefficients relies on the maximum likelihood method. Although the maximum likelihood estimator is a powerful estimation tool, its performance can be severely affected by multicollinearity among explanatory variables. When these variables are correlated, the variances and standard errors inflate, reducing estimator precision. As a result, the maximum likelihood estimator becomes unreliable for estimating regression coefficients. Similarly, in the presence of strong multicollinearity, the Ordinary Least Squares estimator also yields misleading results. In this paper, we propose a new estimator and compare it with existing methods, including the Ordinary Least Squares estimator, Ridge Regression estimator, Liu estimator, and Two-Parameter estimator. The discussion further addresses the estimation of biased parameters and examines the necessary and sufficient conditions associated with the proposed approach. We evaluate the performance of the new estimator using the matrix mean squared error criterion and a simulation study that considers various sample sizes, degrees of multicollinearity, parameter settings, and variance levels. To illustrate practical applicability, we analyze the Theoph dataset, demonstrating that the proposed estimator performs more efficiently than the maximum likelihood estimator and other established techniques. The results show that the generalized two-parameter estimators achieve lower mean squared error than maximum likelihood estimators and deliver superior performance compared to existing methods.
The concept of the odds ratio introduces a powerful measurement that extends the usefulness of odds by incorporating the correlation between two risk factors. Odds ratio, defined as the relative strength of the association between two risks, has the ability to reveal the most influential risk factors associated with a given outcome. This knowledge has the potential to shape preventive strategies and improve public health interventions. Across a diverse spectrum of risk factors and outcomes, odds ratio is an adaptable tool that enriches the arsenal of researchers and public health practitioners. The focal point of this paper is to introduce an analytical expression governing the distribution of the odds ratio in the cases of two independent risks, each following a classical Beta distribution. We provide explicit formulations for the moments of the proposed distribution. This provides a comprehensive understanding of the statistical properties of the distribution. Through simulations under various parameter settings, we elucidate the characteristics of the proposed distribution. Our aim is to showcase its versatility in capturing diverse scenarios. To illustrate its application, we use data from an ongoing cardiovascular study based in Framingham, Massachusetts. The dataset contains 4,238 records and covers 15 demographic, behavioural, and medical risk factors, this study serves as a practical example. Our analysis demonstrates that the proposed distribution appropriately models the relationship between the risk of coronary heart disease and the range of factors involved, including age, sex, education level, and smoking. Therefore, this paper introduces the distribution of the odds ratio as a powerful metric while offering an analytical foundation for its distribution. By demonstrating its efficacy in a substantial cardiovascular study, we underline its practical value in understanding and quantifying the interaction between risk factors and health outcomes.
In large-scale multi-building and multi-floor environments, fingerprint-based indoor localization is challenged by environmental non-stationarity, device heterogeneity, and the cost of maintaining separate models per building and floor. Existing deep methods either rely on unstable three-dimensional coordinate regression or adopt cascaded pipelines that first classify building/floor and then perform local regression, which amplifies upstream errors and increases latency and energy consumption. We propose Multi-Grid Hypothesis Learning (MGHL), a unified framework that performs building-floor recognition and continuous position estimation within a single backbone and a single forward pass. MGHL discretizes the global 2D space into grid cells and uses distance-aware multi-label supervision to construct multiple spatial hypotheses. A robust training objective combining soft-label cross-entropy, focal loss, and label smoothing improves probability calibration and tail performance, while lightweight random masking of access-point (AP) features during training simulates AP failures and temporal drift without adding inference cost. MGHL adopts a shared-embedding multi-head architecture to jointly model building, floor, and grid-level outputs, with a masking mechanism restricting hypotheses to the predicted building-floor subset. Experiments on three benchmark datasets, namely UJIIndoorLoc, SODIndoorLoc, and UTSIndoorLoc, under the EvAAL protocol show that MGHL consistently improves joint building-floor-position accuracy and mean localization error over strong single-model baselines, while approaching the overall performance of traditional multi-model systems.
Vector embeddings make complicated data extracted from networks, words and images, more amendable to data science applications. At the present time, the Veronese-Whitney (VW) matrix embedding of the real projective space is the state of the art for making inference about digital images from an uncalibrated camera, such as a cell phone or security camera. In this work we consider vector embeddings for the projective shape data and in particular determine the minimum dimension isometric (distance-preserving or Nash) vector embedding for a projective space. We determine such an embedding for the projective plane in closed-form. From this embedding we determine an embedding for the Cartesian product of projective planes which is used to develop a novel extrinsic mean test as well as a novel homogeneity test for 2D projective shape analysis. In a Monte Carlo study and real data application it is found that this new testing procedure performs as well as the state of the art extrinsic test based on the VW embedding in terms of hypothesis tests for extrinsic means, tests for homogeneity with tangential components and classification via support vector machines. Furthermore, it generally outperforms the vech of the VW embedding. Note however that the Nash embedding is into five-dimensional Euclidean space, whereas the VW embedding is into the Euclidean space of 3 by 3 symmetric matrices, which is six-dimensional. Our vector-valued Nash embedding is preferred over the matrix-valued VW embedding for data science applications since (i) it is a vector embedding and performs as well as the state of the art VW matrix embedding when the latter can be used in a statistical procedure and (ii) and our embedding is easily used for classification and visualization with traditional statistical techniques.
The rapid advancement in education and technology in various fields has enhanced the students’ future. The forecast of students’ educational performance paves the way for effective learning and training, aiming to improve their learning ability based on the prediction results. Various Deep Learning (DL) models demonstrated their effectiveness in the prediction process by analyzing the student data. The existing methods show their capability in prediction; however, they are not able to handle large datasets. To overcome this, a DL-based hybrid Pyramid Mobile Forward Harmonic Network (PyMFH-Net) model for student performance prediction is proposed. The data transformation is carried out initially by Yeo-Johnson, which makes the dataset more compatible. The essential features are identified through an ensemble feature selection process. Thereafter, the augmentation process enlarges the dataset, and the augmented features are fed into the PyMFH-Net model for the prediction process. The PyMFH-Net model’s performance is validated by using the evaluation metrics and compared with other existing methods. The results demonstrate that the PyMFH-Net model achieved superior performance for dataset 1, with a precision of 91.89
The accurate classification of breast cancer histopathology images is critical for early diagnosis, risk stratification, and treatment planning. However, traditional supervised learning approaches are hindered by the limited availability of annotated medical datasets. At the same time, single-model classifiers often struggle to generalize effectively across subtle morphological variations in tissue structures. To address these challenges, this study proposes CAE-StackNet, a hybrid framework that combines self-supervised feature learning via a Convolutional Autoencoder (CAE) with a heterogeneous stacked-ensemble classifier. The CAE is first trained to reconstruct breast cancer histopathology images from the BACH dataset, enabling it to learn compact, morphology-preserving features without requiring class labels. These self-learned features are then classified using a stacked ensemble of Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP), with a meta-learner aggregating their predictions to produce the final diagnosis. Experimental results demonstrate that CAE-StackNet achieves an AUC of 0.99 and an average precision of 0.96, outperforming individual classifiers trained on the same features. MLP achieved the highest standalone accuracy (0.89), but the stacked ensemble exhibited superior discriminative power, reflecting the advantage of combining complementary classifiers in a diverse ensemble framework. For breast cancer classification in histopathology, CAE-StackNet provides a data-efficient and clinically meaningful solution by reducing the need for manual annotation and improving robustness through ensemble learning. The developed method enables more reproducible and objective histopathological grading and has excellent potential for incorporation into computer-aided diagnostic (CAD) systems. To increase clinical interpretability and gain the trust of pathologists, future efforts will involve external validation on multi-center datasets, integrating pathology-aware constraints into the CAE training process, and creating explainability tools. Recent transformer‑based frameworks, such as DFViT, fuse convolutional and vision‑transformer blocks to capture both local and global tissue patterns and have achieved outstanding accuracy on the BACH and BreakHis datasets. By contrast, CAE‑StackNet achieves comparable AUC and average precision while relying primarily on self‑supervised features and requiring far fewer annotated images, underscoring its potential for scalable clinical deployment.
Probabilistic models have become one of the most powerful statistical methods for modeling random phenomena, particularly the Probabilistic Self-Organizing Map (PrSOM), which is a probabilistic interpretation of the classical Kohonen model (SOM). This algorithm approximates the distribution of data density using a mixture of normal distributions. However, the likelihood function of the normal mixture is unbounded and exhibits some local maxima (degeneracy). In this context, this paper aims to solve this problem. For this purpose, we apply specific constraints to the PrSOM model based on Ingrassia’s approach as well as directly integrate them into each step of the PrSOM algorithm during the phase of updating the covariance matrices. Indeed, this approach gives rise to a new algorithm called a constrained probabilistic self-organizing map (CPrSOM). We give the implementation of our proposed (CPrSOM) method. We then evaluate its effectiveness by comparing its performance with that of other classification methods through numerical simulations and by applying it to a real dataset.
In machine learning, models are typically constructed by learning patterns from user-provided data. When users request the deletion of their data, it is necessary to remove this data’s influence and contributions from the model. However, current research rarely considers the issue from the user’s perspective. As a result, it has not fully explored whether the requested data for deletion truly needs to be removed entirely. In this paper, a novel data description and unlearning framework called Specific Synaptic Dampening over Data Description (SSDD) is proposed. From the perspective of user requests, data description is used to identify which requested data do not require forgetting, and this analysis is further extended to the entire dataset. Next, based on the data description, the correlation between the data and the model is defined. Finally, an adaptive neuron constraint algorithm is designed based on this information to improve forgetting efficiency and model performance. Theoretical analysis and experiments on representative datasets demonstrate the effectiveness and efficiency of SSDD, confirming that not all requested data need to be forgotten in reality.
The research examines foreign value added (FVA) embedded in export-oriented supply chains by integrating public financial disclosures with value-added trade data. Using China as an empirical case, the study applies big data analytics to investigate how the working capital ratio, FVA, and domestic value added (DVA) interact across sectors to influence export competitiveness. Additionally, the study proposes a novel framework that links supply chain finance (SCF) accessibility to disclosure transparency, particularly for small and medium-sized businesses (SMBs). The results indicate that sectors with more advanced disclosure practices and greater digital finance development tend to retain higher levels of domestic financial value and demonstrate stronger financial capacity. A cross-sectoral comparison further reveals significant differences in financial value retention and availability. To enhance the inclusiveness and resilience of global value chains, the findings highlight structural asymmetries in financial value capture and underscore the need for targeted SCF policies, improved standardization of financial indicators, and stronger digital finance infrastructure. This study contributes to the literature by providing a comprehensive perspective on FVA within supply chains. It also offers practical implications for managers, policymakers, and financial service providers seeking to improve financial integration and sustainability in global trade.
Social media serves as a platform to unite individuals with similar interests and perspectives. Twitter, in particular, stands out as a pioneer, significantly influencing people's daily lives by providing a space to express opinions. However, relying solely on hashtags for tweet recommendations might not be the most accurate approach. To enhance this, our focus has been on developing a recommendation system based on both topic and sentiment analysis. Our methodology involved aggregating diverse datasets from sources like Kaggle and merging them into a comprehensive dataset. Employing various models, including LSTM, Logistic Regression, and SVM, we aimed to create a system that enables groups to share their opinions effectively. In our evaluations, the LSTM model showcased remarkable accuracy at 83
Landmarking is a dynamic prediction technique used for analysing time-to-event data with time-dependent covariates. The landmark approach enables us to update the survival probability and the hazard rate of patients, as new covariate information becomes available during the follow-up. These dynamic updations can be made by fitting survival models for individuals who are still at risk at each landmark time point. In this study, we propose parametric proportional hazards (PPH) models to exploit the relation between lifetime and time-dependent covariates, which updates the survival probability and the hazard rate using landmarking approach. We use the Weibull and Gompertz PPH models for dynamic survival prediction. Inference procedures are carried out using maximum likelihood estimation method. We conduct Monte Carlo simulation studies to validate the finite sample behavior of the proposed dynamic prediction model. The practical utility of the procedures is illustrated by applying them to a real dataset on primary biliary cirrhosis disease.
Accurate hourly predictions of particulate matter (PM2.5) concentrations play a critical role in minimizing the harmful effects of atmospheric pollution in China. However, this task is challenged by the non-linear and complex nature of environmental data, often resulting in inaccurate prediction systems. Therefore, to overcome this drawback a novel Multilevel Decomposition and the Dynamic Temporal Fusion Transformer (DTFT) is proposed for accurate prediction of PM2.5 Concentration. The data decomposition includes three stages namely, Multilevel Variational Mode Decomposition (MVMD), Shannon Energy Entropy (SEE) measurement, and Fuzzy C-means (FCM) clustering. MVMD is employed to decompose the time sequence into a set of intrinsic mode functions (IMFs) iteratively, in which every level concentrates on capturing both Fine and coarse frequency components. The SEE measurement is used to evaluate the energy distribution and analyze the importance of IMFs, and FCM clustering groups ensuring the analysis of latent trends and patterns in the atmospheric conditions. The model reconstruction step combines the filtered and decomposed IMFs to reconstruct the real-time series, guaranteeing the reduction of noise and the ensuring of significant signal features. This reconstruction validates the efficiency of the MVMD process and prepares the information for downstream analysis. The DTFT captures long-term dependencies and complex multi-variable interactions to enhance prediction accuracy and robustness. The experimental evaluation shows that the proposed method is superior to other traditional methods. The proposed model attained high R2 of 0.969, 0.989, 0.989, 0.986, and 0.997 in China’s provinces of Guangzhou, Chengdu, Beijing, Chongqing, and Shanghai, respectively.
Diabetes is a chronic metabolic disorder with a growing global burden, making early diagnosis essential for effective management. This study proposes a predictive framework that integrates four machine learning algorithms (MLAs)-gradient boosting (GDB), K-nearest neighbors (KNN), gaussian naïve bayes (GNB), and logistic regression (LGR)–with four feature selection techniques (FSTs), namely genetic algorithm (GA), recursive feature elimination (RFE), Pearson correlation coefficient (PCC), and least absolute shrinkage and selection operator (LASSO). The framework was evaluated on the PIMA Indian Diabetes dataset using standard classification metrics, including accuracy, precision, sensitivity, specificity, F1-score, and the area under curve (AUC). Statistical analysis using the Friedman test confirmed significant performance differences among the feature selection techniques, with PCC achieving the best rank. The optimal results were obtained by combining the GDB classifier with GA and PCC-based feature selection, achieving accuracies of 95.14
Recent recruitment trends are shifting from traditional reactive recruitment toward proactive talent acquisition, particularly in the IT industry, where the demand for highly skilled developers is rapidly growing. While the integration of artificial intelligence (AI) into human resource management (HRM) has reshaped the recruitment landscape, most previous attempts to leverage AI in recruitment have focused on supporting traditional reactive recruitment. This study proposes a novel approach for automating the proactive talent acquisition process for qualified developers. The proposed approach, called LibMatch, automatically produces a list of qualified developers who meet the requirements of hiring companies given a job posting as input. Central to the proposed approach are pre-trained language models, including KeyBERT and SentenceBERT, which are used to extract critical terms from job descriptions and match them to relevant technology libraries that developers have used. The developer candidates are extracted from GitHub, the largest and most popular social coding platform, with more than 100 million developers worldwide. An illustrative example of talent acquisition for natural language processing engineers is presented to demonstrate and validate the proposed approach. The proposed approach enables recruiters to efficiently and effectively identify qualified candidates from a talent ocean without any additional effort beyond entering job postings. The code is available at https://github.com/Seoultech-Inno/LibMatch .
In this paper, a Hierarchical Cross-Modal Alignment Optimization Framework (HCAOF) is proposed with the intention of enhancing the robustness and consistency of semantic relationships in multimodal translation. The model incorporates a Bidirectional Adversarial Alignment Mechanism (BAAM) for fine-grained visual-textual alignment, a Dynamic Semantic Enhancement Network (DSEN) for noise filtering and spatial modeling, and a reprojection gradient optimization for managing long-tail and weakly aligned samples. All of these components work together to achieve the desired results. The approach was evaluated using the MS-COCO, Flickr30K, and TextILE datasets, and it achieved a BLEU-4 score of 38.3 and a METEOR score of 33.2. This marked a 2.4
In the intricate landscape of dermatological diagnostics, traditional methodologies often necessitate the specialized expertise of healthcare professionals to discern between subtly varying skin conditions. To bridge this gap, our research advances a computational framework that amalgamates classical convolutional neural networks (CNNs) with custom-designed layers for nuanced skin lesion classification. The model accommodates 224 × 224 pixel images across three color channels (RGB) and counters inconsistencies such as variable lighting through data augmentation techniques like random contrast adjustments. Established architectures like VGG16 and ResNet101 are employed via transfer learning for initial feature extraction. Unique to our model are custom layers: a Fourier layer for frequency domain characteristics, a Laplacian layer for edge detection, a Gabor layer for texture analysis, an RGB-to-HSV layer for color variation, and a histogram equalization layer for image contrast. These are aggregated in a ‘CombinedOutput’ layer and processed by feed-forward layer-X and layer-A using GELU activation and batch normalization. A final concatenation layer leads to a softmax-activated prediction layer. We employed exhaustive performance evaluation metrics: Accuracy, Precision, Recall, AUC, F1. Our best-performing custom model exhibited a Recall of 94.2
We present a method for analyzing the decision boundary in deep learning models, focusing on the study of the minimum distance of images to this boundary and the evolution of this margin throughout the training of a convolutional neural network. To enhance the robustness of the model, we generate, in addition to the original images, a perturbed version of each image while preserving the same label. The dataset used includes 15 types of otoliths, calcified structures located in the inner ear of fish. Recognizing their shape provides economically and ecologically relevant information about the status of fish populations worldwide. Indeed, otolith analysis allows researchers to determine species stocks, estimate fish age, and study taxonomy. However, identifying species from otolith images remains a labor-intensive task requiring specialized expertise and considerable resources. Recent approaches to otolith image classification mainly rely on deep neural networks, whose primary challenge is generalizability – the ability to maintain high performance on unseen data. Our study aims to better understand this issue by evaluating the decision boundary throughout the training process. Our results show that the convolutional neural network we developed achieves an accuracy of 92
