Debabrata Pruseth

Explainable Artificial Intelligence (XAI) is becoming essential for organizations that rely on machine learning for critical decision-making. This research presents a practical framework for interpreting and trusting machine learning models using complementary explainability techniques, including SHAP, LIME, ELI5, DALEX, Partial Dependence Plots, and Individual Conditional Expectation analysis. Through a decision-tree income-classification case study, the research demonstrates how global and local explanations can reveal model behavior, feature importance, decision pathways, and prediction logic. The findings show that no single explainability method is sufficient on its own; instead, trustworthy AI requires a combination of interpretability techniques, governance practices, continuous validation, and stakeholder-focused communication. The proposed framework provides a practical roadmap for implementing transparent, accountable, and reliable machine learning systems in real-world environments.

Machine learning systems increasingly influence critical decisions affecting individuals and communities. While models may demonstrate strong overall performance, hidden bias can produce unequal outcomes across demographic groups, leading to fairness concerns and regulatory risks. This study introduces a structured framework for identifying protected attributes, evaluating group-level outcomes, measuring disparate impact, and applying practical bias mitigation strategies. The approach integrates technical fairness assessment with governance principles such as transparency, accountability, auditability, and continuous monitoring. Designed for practitioners, researchers, and decision-makers, the framework provides actionable guidance for improving fairness in AI systems used in employment, healthcare, education, lending, insurance, and public-sector applications while supporting responsible and trustworthy AI deployment.

Time series forecasting plays a critical role in modern machine learning applications across finance, healthcare, retail, manufacturing, energy, and cybersecurity. This article presents a practical framework for understanding and implementing time series modeling techniques, from classical statistical approaches such as ARIMA and exponential smoothing to advanced deep learning architectures and emerging foundation models. The framework covers forecasting, anomaly detection, classification, clustering, segmentation, and similarity search while addressing preprocessing, feature engineering, temporal validation, and model interpretability. Researchers and practitioners can use this structured methodology to select appropriate forecasting strategies and build robust data-driven temporal analysis solutions.

Transfer learning has transformed natural language processing by enabling pretrained language models to be adapted efficiently for downstream tasks. This paper provides a practical framework covering feature extraction, full and partial fine-tuning, adapter-based learning, prompt tuning, instruction tuning, and zero-shot inference. It discusses leading transformer architectures including BERT, RoBERTa, DistilBERT, BioBERT, SciBERT, GPT, and T5 while outlining implementation strategies such as gradual unfreezing, discriminative learning rates, domain-adaptive pretraining, and parameter-efficient methods like LoRA and BitFit. The framework helps practitioners evaluate trade-offs among accuracy, computational cost, data requirements, deployment constraints, and governance considerations when selecting transfer learning approaches for real-world NLP systems.