Journal of Applied Statistics and Machine Learning

There has been a comparative analysis done on fifteen classifiers (Models). When it came to predicting the students’ admission status (binary), the Random Forest classifier (RF) was ranked No 1. Metrics including Accuracy (95.5%), Precision (97.2), Recalls (93.2) and F1-Score (95.2) were used to assess the models’ performance on test/validated data. Accuracy (95.5%), Precision (95.3), Recalls (94,9) and F1 (95.0). The area under the ROC curve was also very high (0.99), suggesting that the RF model is the most effective and perfect model and that it can be used to predict the admission status for new applicants to educational Institutions.

This study fitted the consumer price index for industrial workers (CPIIW) in India using a time series data on Consumer Price Index for Industrial Workers. A comparative evaluation of 28 Machine learning Models for time series data has been carried out . An invariant of Recurrent Neural Network Model viz. Long Short-Term Memory Model(LSTM) was found to be best fit to the CPIIW data .The Performance of the Models were assessed by the four matrices : Mean Absolute Error (MAE), Root Mean Square Error (RMSE) ,Mean Absolute Percent Error (MAPE) and R Square (R2). It offered a two-year projection for the anticipated CPIIW. The CPIIW two-year forecast, which runs from January 2024 to December 2025, was finally made. According to the study, the CPIIW will increase by about 3.4 points during the summer/rainy season (April to September) and stay relatively stable during the winter (October to March).

Keywords: LSTM, CNN, RNN, DEEP LEARNING, RMSR, MAPE, MAE, R2

Advances in research technologies make high-dimensional data available, and the most interesting research has been conducted on variable selection. The Bayesian variable selection is gaining interest in different fields because of its vast literature. In this paper our main objective is to investigate Bayesian variable selection (BVS) in the context of a random effect model using various prior for variance components in the spike-slab prior approach.

The spike-slab prior is viewed as a mixture distribution, where some regression coefficients concentrate around zero (spike), and the remaining coefficients have a probability of not being zero values (slab). Generally, for building a model, the normal distribution is considered a prior for regression coefficients in the spike-slab components. When the variance is unknown in normal, it could be estimated by extending the model into a hierarchical model. In a hierarchical model, different prior distributions have been proposed for the variance components. Generally, the prior distributions chosen for the variance component are uniform, inverse gamma, and half-Cauchy distribution. Both simulation and real data will be studied to investigate and evaluate how good the chosen distribution for variance components is in the random effect model for the spike-slab approach.

The application of the BVS with spike-slab prior to microarray data from ADNI (Alzheimer’s Disease Neuroimaging Initiative), in a logistic regression setting (Alzheimer’s Disease vs. Control) demonstrated a notable degree of dimensionality reduction. These selected genes maintain lower misclassification error percentages with higher area under the receiver operating characteristic curve( AUC-ROC) values in different machine-learning algorithms. This discovery opens up new avenues for in-depth exploration and investigation, potentially leading to the identification of biomarkers for Alzheimer’s Disease(AD).

The COVID-19 pandemic which began as an epidemic demonstrates the need for a better system and technologies to monitor and identify health emergencies before they become a disaster that claims human lives and damages the economy. The twenty-first century is known for technological advancements, particularly in the fields of Artificial Intelligence (AI) and Machine Learning (ML). A number of AI and ML-based algorithms and models have been developed to date for surveillance and precision decision-making in the healthcare domain. Machine Learning, as represented by Supervised Machine Learning, Unsupervised Machine Learning, Reinforcement Learning, and Semi-Supervised Machine Learning algorithms, has made considerable progress in the field of health care, However, there is still room for advancement. The purpose of this review paper is to identify the models developed for epidemic assessment and prediction, and simultaneously identify which areas of the healthcare system need improvement and how Machine Learning models can help.

Keywords: Pandemic, Epidemic, Artificial Intelligence, Machine Learning, Surveillance and Decision-making.

Artificial intelligence (AI) has revolutionized customer support by improving efficiency and enhancing user experiences. However, traditional machine learning (ML) methods often rely on extensive local training with sensitive data, raising significant privacy concerns and posing challenges for compliance with regulations like the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA). While privacy-preserving methods such as anonymization, differential privacy, and federated learning address some of these concerns, they often struggle with scalability, utility, and implementation complexity.

This paper presents the Privacy-Preserving Zero-Shot Learning (PP-ZSL) framework—a groundbreaking approach that harnesses the power of large language models (LLMs) in zero-shot learning. Unlike traditional ML techniques, PP-ZSL avoids local data training by leveraging pre-trained LLMs to generate accurate responses directly. The framework integrates real-time data anonymization to protect sensitive information, retrieval-augmented generation (RAG) for handling domain-specific queries, and advanced post-processing mechanisms to ensure compliance with privacy regulations. Together, these elements reduce risks, simplify compliance, and enhance scalability and efficiency.

Our analysis reveals that the PP-ZSL framework delivers privacy-compliant, accurate responses while minimizing the cost and complexity of deploying AI-driven customer support systems. The framework holds transformative potential for industries such as financial services, healthcare, e-commerce, legal support, telecommunications, and government services. By balancing privacy with performance, PP-ZSL sets the stage for secure, efficient, and regulation-ready AI solutions in customer interactions.

Keywords: Privacy-Preserving Zero-Shot Learning (PP-ZSL), Large Language Models (LLMs), Data Privacy and Anonymization, Zero-Shot Learning (ZSL), Regulatory Compliance (GDPR, CCPA), Retrieval-Augmented Generation (RAG).