Trustworthy Dimensionality Reduction

TL;DR

This paper introduces LSDR, a new dimensionality reduction method that balances trustability and generalizability, outperforming existing algorithms like tSNE and UMAP in preserving data structure.

Contribution

It formally models trustability and generalizability in dimensionality reduction and proposes LSDR, an algorithm that optimally balances these aspects, with extensions for broader applicability.

Findings

LSDR outperforms tSNE and UMAP in global structure preservation.

LSDR effectively balances local detail and overall data integrity.

Proposed indices measure trustability and generalizability in dimensionality reduction.

Abstract

Different unsupervised models for dimensionality reduction like PCA, LLE, Shannon's mapping, tSNE, UMAP, etc. work on different principles, hence, they are difficult to compare on the same ground. Although they are usually good for visualisation purposes, they can produce spurious patterns that are not present in the original data, losing its trustability (or credibility). On the other hand, information about some response variable (or knowledge of class labels) allows us to do supervised dimensionality reduction such as SIR, SAVE, etc. which work to reduce the data dimension without hampering its ability to explain the particular response at hand. Therefore, the reduced dataset cannot be used to further analyze its relationship with some other kind of responses, i.e., it loses its generalizability. To make a better dimensionality reduction algorithm with a better balance between these two, we shall formally describe the mathematical model used by dimensionality reduction algorithms and provide two indices to measure these intuitive concepts such as trustability and generalizability. Then, we propose a Localized Skeletonization and Dimensionality Reduction (LSDR) algorithm which approximately achieves optimality in both these indices to some extent. The proposed algorithm has been compared with state-of-the-art algorithms such as tSNE and UMAP and is found to be better overall in preserving global structure while retaining useful local information as well. We also propose some of the possible extensions of LSDR which could make this algorithm universally applicable for various types of data similar to tSNE and UMAP.