A COMPARATIVE STUDY OF DIMENSIONALITY REDUCTION TECHNIQUES FOR HIGH-DIMENSIONAL STATISTICAL DATA

• Dhuha Salim Waheed

  AL-Furat AL-AwsatTechnical University, Al-Qadisiyah Polytechnic College, Iraq

  Author

• Zahraa Saad Jasim

  AL-Furat AL-AwsatTechnical University, Al-Qadisiyah Polytechnic College, Iraq

  Author

• Mohammed Guraibawi

  AL-Furat AL-AwsatTechnical University, Al-Qadisiyah Polytechnic College, Iraq

  Author

Dimensionality reduction is: a necessary processing step in order to properly analyze large data sets with many variables; makes it easier to visualize data structures, and reduces computational complexity; reduces the curse of dimensionality. Three popular techniques for reducing dimensionality in high-dimensional datasets were compared with one another for this study. They are: Principal Component Analysis; t-Distributed Stochastic Neighbor Embedding (t-SNE); and Uniform Manifold Approximation and Projection (UMAP). The data used here is derived from the classic Iris dataset augmented by 50 random features obtained through some other means. According to PCA, linear projections can be used while still retaining maximum variance. t-SNE and UMAP give non-linear representations that allow for both local and global structure. Our experiments show that all methods preserve the underlying class structure, while t-SNE and UMAP provide more sharply clustered results. Silhouette analysis confirms the quality of clusters. These results indicate a trade-off between linear and non-linear methods to reduce dimensionality in high-dimensional data.

• pdf

This work is licensed under a Creative Commons Attribution 4.0 International License.