The main goal of dimension reduction methods is to make the dimension of the transformed representation correspond with the internal dimension of the data. In other words, it should be similar to the minimum number of variables necessary to express all the possible properties of the data. Reducing the dimension helps mitigate the impact of the curse of dimensionality and other undesirable properties that occur in high-dimensional spaces. As a result, reducing dimensionality can effectively solve problems regarding classification, visualization, and compressing high-dimensional data. It makes sense to apply dimensionality reduction only when particular data is redundant; otherwise, we can lose important information. In other words, if we are able to solve the problem using data of smaller dimensions with the same level of efficiency and accuracy, then some of our data is redundant. Dimensionality reduction allows us to reduce the time and computational costs of solving a problem. It also makes data and the results of data analysis easier to interpret.

It makes sense to reduce the number of features when the information that can be used to solve the problem at hand qualitatively is contained in a specific subset of features. Non-informative features are a source of additional noise and affect the accuracy of the model parameter's estimation. In addition, datasets with a large number of features can contain groups of correlated variables. The presence of such feature groups leads to the duplication of information, which may distort the model's results and affect how well it estimates the values of its parameters.

The methods surrounding dimensionality reduction are mainly unsupervised because we don't know which features or variables can be excluded from the original dataset without losing the most crucial information.

Dimensionality reduction methods can be classified into two groups: feature selection and the creation of new low-dimensional features. These methods can then be subdivided into linear and non-linear approaches, depending on the nature of the data and the mathematical apparatus being used.