A credit card company wants to speed up training of a credit scoring model built from thousands of raw attributes, many of which are highly correlated. Training is slow and the model overfits. Which feature-engineering technique will reduce dimensionality and training time without discarding much information?

Using an autoencoder or Principal Component Analysis (PCA) is the most effective technique here because both methods are designed for dimensionality reduction while preserving the most important information. PCA transforms the original features into a smaller set of uncorrelated principal components, which are linear combinations of the original features. Autoencoders learn a compressed, lower-dimensional representation of the input data in an unsupervised manner. Both techniques address the issues of high correlation, slow training, and overfitting by reducing the number of features without significant information loss. Computing feature self-correlation and dropping highly correlated features is a simpler approach but can lead to information loss if the dropped features still contain unique predictive power. Normalizing features scales them but does not reduce dimensionality or address correlation. Clustering data with k-means and sampling points creates a smaller dataset but doesn't reduce the number of features or address the correlation between them; it reduces the number of data points, which is not the primary problem described.