Dimensionality Reduction in Machine Learning


Welcome readers !! This is the last article on theoretical study of machine learning.

It is computationally expensive and time-consuming for a machine to perform the operations with large datasets. It is because more images implies more features. So, to overcome this issue, dimensionality reduction is performed. These are assumption-based methodologies.

Most commonly used dimensionality reduction techniques are principle component analysis (PCA), factor analysis (FA), and non-negative factor minimization (NFM).

Principle Component Analysis

This methodology is based on the assumption that first few components of the dataset carries most variance. First component has highest variance, which is followed by second component, and so on. These components are transformed linearly to generate new components. Thus, it can be said that all the primary components are the linear combination of the original components. When all the principle components become orthogonal to each other, no dimensional redundancy is performed further.

Eg: Engine Health Monitoring


Factor Analysis

The learning of any machine might be affected by some unobserved or latent or common factors. Each component is assumed to be dependent on the combination of these unobserved factors.

Eg: Tracking Stock Price Variation


Non-negative Matrix Factorization

This methodology is used to represent the features in a product of matrices such that all the matrices have only non-negative elements. It is assumed that the approximate number of features might be reduced.

Eg: Text Mining


Reference: Machine Learning with MATLAB (eBook)


Akshay Juneja authored 10+ articles for INFO4EEE Website on Deep Learning.

No comments

Powered by Blogger.