A machine learning engineer preparing data for a classification model notices a few continuous numeric features have values much larger than the others. A domain expert confirms the features are informative and the dataset is representative. The trained model's accuracy is lower than expected. Which preprocessing step will most improve inference accuracy?

Normalizing features with disproportionately large values (e.g., using Min-Max scaling or Z-score standardization) helps prevent these features from dominating the learning process, especially for algorithms sensitive to feature scales like SVMs, neural networks, or K-nearest neighbors. When features have vastly different scales, the model might incorrectly prioritize the larger-valued features, leading to suboptimal performance and lower accuracy. Bootstrapping is a resampling technique for estimating statistics, not for addressing scale differences. Removing informative features would lead to a loss of valuable information. Extrapolating synthetic features from existing ones is generally not a standard or beneficial preprocessing step for scale issues and could introduce noise or bias.