Traditional feature selection methods focus on reducing dimensionality, often at the cost of discarding valuable but hidden information. In contrast, this paper introduces Feature Expansion, a novel approach that dynamically identifies and reconstructs missing features, enhancing predictive performance without requiring additional data collection. We propose the Semi-Dynamic Feature Set (SDFS) Network, a unique machine learning framework that learns both feature values and model parameters, thereby improving robustness in complex datasets. To enhance interpretability, we integrate Large Language Models (LLMs) to assign meaningful labels to the newly discovered features, transforming them from black-box variables into actionable insights. Our proposed feature expansion method consistently improved classification performance across multiple datasets. In particular, the accuracy increased from 89% to 95% for Parkinson's Disease Detection and from 82% to 85% for Wine Quality Prediction, demonstrating its effectiveness in discovering missing features. By combining feature expansion with LLM-driven explainability, this method not only enhances model accuracy but also unveils hidden data patterns in critical domains such as healthcare, finance, and scientific research. The full implementation is available on GitHub [GitHub link], and the model can be installed via PyPI under the package 'sdfs' [PyPI link].