Dimension reduction, also called dimensionality reduction, is the process of transforming data described by many variables into a representation with fewer variables while retaining the structure or variance that matters. Techniques include principal component analysis, t-SNE, UMAP, and autoencoders. Analysts use it to lower storage and computation costs, reduce noise, counter the curse of dimensionality that degrades distance-based algorithms, and enable two- or three-dimensional visualization of complex datasets. The qualifier in the question, preserving meaningful attributes of the original data, is the signature of this concept.
Regression is a supervised learning task that estimates a continuous target from input features, such as predicting house prices. It maps inputs to an output value and does not alter how many dimensions the data has.
Dimension enhancement is not an established term in machine learning. The nearest real idea is feature engineering, where new variables are created or expanded, but that moves in the opposite direction from the described operation and has no standard name by which it would appear in a curriculum.
Since the description specifies conversion from high-dimensional to lower-dimensional form with information preserved, dimension reduction is the exact fit. It is frequently applied during data preparation, before clustering or classification.
[References:AI Concepts: Data Preparation, Feature Engineering, and Dimension Reduction.]
Submit