Enhancing t-SNE with PCA Initialization and Hybrid Distance Metrics

Wadkar, Srushti Shrikant (2025) Enhancing t-SNE with PCA Initialization and Hybrid Distance Metrics. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

High-dimensional data is often employed in the modern data analysis, but the complexity of this data type impairs visualization and interpretation. The t-distributed Stochastic Neighbour Embedding (t-SNE) algorithm is a common method of dimensionality reduction and particularly effective at preserving local data structure of 2D or 3D projections. However, there are two important limitations of t-SNE: finite memory of the global structure that can cause biases in the interactions within clusters and the inability to reproduce results because of random initialisation.

When solving these limitations, this study has two specific contributions to make PCA-based initialization and a hybrid distance measure, which can be expressed as a combination of cosine similarity and Euclidean distance. Initialization of PCA also gives an initial state that t-SNE optimization can rely on, to improve repeatability and preserve the global variance structure. So as to improve the stability and interpretability of clusters, the hybrid distance measure is a trade-off between angular and magnitude similarity.

Having been implemented in Python with the help of scikit-learn, the techniques were tried out on a range of datasets, including MNIST image set and financial records of bankruptcies. The embeddings were visually inspected in a qualitative way, and silhouette scores were taken in a quantitative way. The results showed that the hybrid distance measure performed better in yielding globally consistent embeddings particularly when a complex data is under consideration, but initialization by PCA was more consistent and reproducible in its layouts.

The paper confirms and scales up the results of the earlier research by systematizing and testing these advancements to a wide variety of data domains, presenting practical, repeatable procedures to more effectively visualize high-dimensional data. The results can be useful to researchers and practitioners seeking useful low-dimensional representations of complex data that are easy to understand.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Horn, Christian UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QA Mathematics > Algebra > Algorithms > Computer algorithms
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	27 Aug 2026 09:23
Last Modified:	27 Aug 2026 09:23
URI:	https://norma.ncirl.ie/id/eprint/9682

Actions (login required)

View Item