You have a tabular dataset with 10,000 rows and 1,020 features. There are no missing values, few duplicates, and about 200 feature pairs with correlation > 0.9. Feature means are similar to medians. Which feature-engineering approach should you apply when training a model in SageMaker?

PCA is the most suitable approach here because it effectively addresses the high dimensionality (1020 features) and multicollinearity (200 feature pairs with correlation 0.9) present in the dataset. PCA transforms the original features into a smaller set of uncorrelated principal components, preserving most of the variance while reducing noise and computational complexity. Dropping features with low correlation scores is less effective because it doesn't address multicollinearity among the remaining features. Anomaly detection is irrelevant to the problem of high dimensionality and multicollinearity. Concatenating highly correlated features would exacerbate multicollinearity and not reduce dimensionality effectively.