Exploring Explainable AI Methods for Decoding Machine Learning Model Decisions
Shruti Kumar, Namrata Dhanda, Kapil Kumar Gupta · 2025
ML has progressed at such high speeds that some models now exhibit opaque decision-making, functioning without a clear or logical explanation for their decision-making procedures. This presents a serious problem for transparent systems such as in medical services, financial sectors, and self-driving technologies, where the predictions of model need to be trusted and understood for regulatory compliance. XAI techniques work on providing some level of transparency by explaining the forecast made by ML models, enabling effective stakeholder validation. This paper analyzes different XAI methods, focusing on both architecture-independent approaches and those tailored to specific model types. Model agnostic techniques like LIME and SHAP offers explanations by simulating model performance without changing the system's structure. However, model-specific methods, including decision trees, attention, and even feature visualization in neural networks, offer an intrinsic form of interpretability as they embed the ability to explain the model within the model itself. We also consider rule-based approaches, counterfactual explanations, and gradient attributed methods and their pros and cons in various ML scenarios. Moreover, we consider the balance between interpretability and the model's accuracy, that XAI's contribution to practice, and ongoing research aimed at the enhancement of transparency of deep learning models. In addition, ethical issues like fairness, bias mitigation, or human-centric AI development are also discussed to highlight the importance of responsible AI practice. This study intends to facilitate researchers and practitioners in choosing relevant interpretability techniques through reviewing various XAI techniques. This study serves to contribute to the inavaluable knowledge and research conducted within the XAI arena, paying close attention to building reliable and transparent machine learning models. The objective of this paper is to analyze and compare various Explainable AI methods like SHAP and LIME in terms of their interpretability, computational complexity, and usefulness in real-world decisionmaking scenarios.