Journal of Data Science and Modeling

Abstract Accurate weather prediction plays a vital role in many sectors, such as agriculture, disaster preparedness, transportation systems, and urban planning. Traditional meteorological models face challenges in capturing complex atmospheric dynamics, leading to increased reliance on artificial neural networks (ANNs) for improved forecasting accuracy. ANNs have been widely applied in meteorology due to their ability to model nonlinear relationships and temporal dependencies. Based on the Sinc numerical methods, the modified Sinc neural network (MSNN) has been introduced recently. This model uses the advantages of the Sinc function, such as smoothness and fluctuation, and at the same time improves the ability to model nonlinear dependencies and temporal dynamics in environmental data. This work utilizes the MSNN for time series forecasting where its parameters are adjusted with a discrete-time online Lyapunov-based learning algorithm. Then, it is applied to enhance the weather forecasting. This model is evaluated on datasets containing various meteorological variables. The data used in this article is related to the city of Khorramabad in Iran. The results show that despite its simple structure, MSNN has a high efficiency in weather forecasting.

Abstract Principal Component Analysis (PCA) is a cornerstone technique for dimensionality
reduction and data analysis. However, classic PCA can exhibit instability in
high-dimensional settings where the number of variables significantly exceeds the
number of observations. Shrinkage-based PCA addresses this limitation by incorporating
regularization into the covariance matrix estimation process, leading to
more stable and interpretable results. This paper provides a robust mathematical
and statistical foundation for shrinkage-based PCA, compares its performance with
classic PCA, and demonstrates its advantages through theoretical analysis, numerical
simulations, and real-world data experiments. It is important to note that using the idea of a contraction estimator increases the efficiency of the estimator. mean time in this paper, it is shown that the covariance matrix estimator resulting from the contraction estimator is very efficient.
It is also worth mentioning that to increase the efficiency of the contraction estimator, the recently discussed interval contraction estimator can be used.
keywords: principal component analysis, Shrinkage-based, Estimation, Covariance Structures, Simulation.

Abstract Capillary action and water uptake are tremendously fundamental and practical phenomena which used in a wide range of applications from industries and medical to agricultural fields. This work aims to provide a detailed numerical investigation and statistical data sampling of capillary action and water uptake, considering hysteresis associated with density, surface tension, contact angle, gravity force, tube diameter, and inclination angle effects. The main selected domain is 1mm and 5mm in diameter and height. The solver type is chosen as a pressure-based solver, and time-dependent data sampling is utilized. The flow field is selected as incompressible, constant properties, Newtonian homogeneous fluid. The finite volume method on a co-located grid system is used. The code uses algebraic multigrid schemes to accelerate the solution. The message passing interface parallelized code is used. The bisection algorithms are used for partitioning. The pressure and velocity fields were coupled using the PISO algorithm. The results show that increasing capillary tube diameter or surface tension enhances uptake velocity by 98–100% and reduces filling time by 49–50%, respectively, though inertial/dissipative effects caused minor deviations (1–12%) in surface tension cases. Flow velocity scaled linearly with contact angle doubling filling time, while gravitational acceleration induced only marginal delays with negligible meniscus impact, supporting its omission in engineering models. Transient meniscus asymmetry occurred in inclined tubes (45°) due to contact angle disparity between halves, yet filling duration remained identical to vertical and horizontal orientations despite geometric differences in meniscus evolution.

Abstract Abstract of paper: This study explores the application of Recurrent Neural Networks (RNNs) for predicting loan defaults, with a particular emphasis on incorporating uncertainty estimation into the predictive framework. Conventional RNN models demonstrate high accuracy, but they fail to provide quantitative measures of prediction uncertainty. To address this limitation, a dual modeling approach is proposed: a standard RNN model for achieving high predictive accuracy and an uncertainty-aware RNN model incorporating Bayesian inference. The uncertainty-aware model enables enhanced risk assessment through confidence level estimation and improved capture of complex temporal dependencies in financial data. Experimental results indicate that both proposed models outperform traditional methods, with the uncertainty-aware variant offering superior risk evaluation capabilities through its probabilistic outputs. These findings contribute to advancing credit risk assessment methodologies and offer practical value for financial institutions seeking more robust default prediction systems.
Keyword: Recurrent Neural Networks (RNNs), Loan Default Prediction, Uncertainty Quantification, Credit Risk Assessment‎.

Abstract This article focuses on diagnostic measures for identifying high-leverage points, and influential observations in linear mixed measurement error (LMME) models. It achieves by imposing the stochastic restrictions on the parameters and incorporating the ridge estimator to tackle the issue of multicollinearity. To this end, generalized leverage matrices are defined using the restricted ridge estimator (RRE) to identify high-leverage observations. Additionally, analogs of Cook’s distance and likelihood distance are proposed to determine influential observations through a case deletion approach. Simulation studies and real-life applications support the theoretical results. To the best of our knowledge, there has been no significant attention in the literature regarding diagnostics for leverage and influence measures concerning the outcomes of the RRE in LMME models. Hence, this paper evaluates the influence of observations by using leverage and influence measures to identify influential observations on the RRE’s of fixed effects and the prediction of random effects in LMME models

Abstract In this paper, we propose a new two-parameter discrete distribution based on central Bell expansion, which is zero-inflated and designed to effectively model overdispersed count data. We study several structural properties of the proposed distribution and demonstrate that it is infinitely divisible, which adds theoretical strength and potential for wider applicability. The paper also discusses parameter estimation techniques for the distribution, focusing on two common approaches: the method of moments and the maximum likelihood estimation method. Both methods are developed and explained in detail. To evaluate the accuracy and reliability of these estimators, a simulation study is conducted across different sample sizes, allowing us to assess their performance under various conditions. To illustrate the practical importance and usefulness of the new distribution, we apply it to two real data sets and show how well it fits the observed data, reinforcing its value as a flexible tool for analyzing count data.

Abstract The paper suggests a novel model defined on the unit interval, termed the unit linear exponential distribution, constructed via an inversion of the exponential function. The proposed model contains the unit exponential and the unit Rayleigh distributions as special submodels. Fundamental properties of the introduced distribution are discussed which are stochastic ordering property, quantile function, incomplete moments, moments, probability weighted moment, order statistics, stochastic orderings, stress strength reliability, and Tsallis and Renyi entropies. The distribution has two unknown parameters, which are estimated utilizing the following methods: maximum likelihood, maximum product spacing, least and weighted least squares, Cramér-von Mises, and Anderson-Darling. The behavior of these estimators is assessed through a simulation study. Furthermore, the paper develops a novel quantile regression model based on suggested distribution, which is shown to be a good alternative to existing models like the Kumaraswamy, beta, and unit Chen quantile regression models. We estimate the parameters of the regression model utilizing maximum likelihood. Two well-known real data applications are given to prove the modeling capability of the newly suggested distribution and quantile regression model.

Abstract In this study, a multi-stage approach based on online exam data analysis was proposed to identify and rank students' Cheating Ranks. Initially, student response sheets were clustered using the $K-means++$ algorithm with dynamic determination of K, forming groups with similar characteristics. Subsequently, in later stages, each student's Cheating Rank was determined based on various behavioral and performance parameters.

Results demonstrated that the Cheating Rank derived from online exam data effectively differentiates between students suspicious of cheating and normal students, with statistically significant differences between the two groups. These findings underscore the validity and efficacy of the proposed method in detecting cheating in online exams.

Additionally, the impact of threshold selection for group differentiation highlighted the importance of appropriate parameter tuning in enhancing detection accuracy and influencing model sensitivity and reliability. The use of in-person exam scores as a reference criterion strengthened the results' credibility and enabled more objective model evaluation.

Given the limitations in sample size and data scope, future research should focus on larger, more diverse, and multidimensional datasets to improve both diagnostic accuracy and model generalizability. Furthermore, integrating this approach with advanced machine learning techniques and behavioral analytics could significantly enhance online exam integrity monitoring systems.

Overall, this research represents a significant step toward developing cost-effective, reliable, and efficient methods to reduce cheating in online educational environments, fostering greater trust among instructors and students in assessment processes.

Abstract We investigate the problem of weight uncertainty originally proposed by [Blundell et al. (2015). Weight uncertainty in neural networks. In International conference on machine learning, 1613-1622, PMLR.] in the context of neural networks designed for regression tasks, and we extend their framework by incorporating variance uncertainty into the model. Our analysis demonstrates that explicitly modeling uncertainty in the variance parameter can significantly enhance the predictive performance of Bayesian neural networks. By considering a full posterior distribution over the variance, the model achieves improved generalization compared to approaches that treat variance as fixed or deterministic. We evaluate the generalization capability of our proposed approach through a function approximation example and further validate it on the riboflavin genetic dataset. Our exploration encompasses both fully connected dense networks and dropout neural networks, employing Gaussian and spike-and-slab priors respectively for the network weights, providing a comprehensive assessment of how variance uncertainty affects model performance across different architectural choices.

Abstract Critical services increasingly rely on distributed computing infrastructures that can deliver fast and reliable responses under dynamic conditions. Traditional cloud-centric deployments often struggle to meet these demands due to latency and reliability constraints. This paper addresses these challenges by exploring intelligent service placement across the cloud–edge computing continuum. We introduce an adaptive placement framework that continuously adjusts deployment decisions in response to changing system states and service requirements. To improve robustness, the framework integrates learning transfer mechanisms and a criticality-aware resilience strategy that prioritizes service continuity during failures. The proposed approach enhances responsiveness and reliability, leading to a higher proportion of services meeting strict timing constraints. Comprehensive experimental evaluations confirm that the framework provides more effective support for critical and delay-sensitive services compared to existing placement strategies in continuum-based environments.

Abstract The Weighted Marshall-Olkin Bivariate Exponential (WMOBE) distribution was first proposed by
Jamalizadeh and Kundu (2013), who examined its different characteristics and properties. Bayesian
estimation of the model parameters is carried out using both the squared error loss (SEL) function,
which is symmetric, and the linear-exponential (LINEX) loss function, which is asymmetric. These
estimators are derived under both informative and non-informative gamma priors. Given the complexity
of the four-parameters model, explicit analytical solutions for the Bayesian estimators are not attainable,
making it necessary to employ the Gibbs sampling procedure. Markov Chain Monte Carlo (MCMC)
methods are widely utilized to compute and implement these estimates. Furthermore, the convergence
behavior of the Markov chain toward a stationary distribution is carefully analyzed. Credible intervals,
particularly the highest posterior density (HPD) intervals for the unknown parameters, are also
constructed. To assess and compare the effectiveness of these estimation approaches, Monte Carlo simulations are performed. Finally, the methodology is applied to a real-world dataset for illustrative purposes.

Abstract This study investigates the application of Bayesian two-sample testing in Hilbert space to monitor the health status of industrial equipment. The proposed method evaluates distributional differences between operational data samples to detect faults or anomalies. Unlike traditional multivariate techniques, our approach pro-vides higher sensitivity to subtle distributional shifts and supports visual insights into posterior distributions. Real-world experiments on industrial sensor outputs demonstrate the method’s effectiveness in early fault detection, reducing mainte-nance costs and downtime. This makes Bayesian two-sample testing in Hilbert space a powerful tool for predictive maintenance strategies.
This study investigates the application of Bayesian two-sample testing in Hilbert space to monitor the health status of industrial equipment. The proposed method evaluates distributional differences between operational data samples to detect faults or anomalies. Unlike traditional multivariate techniques, our approach pro-vides higher sensitivity to subtle distributional shifts and supports visual insights into posterior distributions. Real-world experiments on industrial sensor outputs demonstrate the method’s effectiveness in early fault detection, reducing mainte-nance costs and downtime. This makes Bayesian two-sample testing in Hilbert space a powerful tool for predictive maintenance strategies.