Effects of sparseness and randomness of pairwise distance matrix on t-SNE results

Eli Parviainen*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review

1 Citation (Scopus)


We apply ideas from random graph theory to sparse pairwise distance matrices in dimension reduction. We use matrices with some short and some randomly chosen distances, and study effects of matrix sparseness and randomness on trustworthiness and continuity of t-SNE visualizations. The existing works have either concentrated on matrices with only short distances, or implemented heuristics with mixed distances without explaining the effects. We find that trustworthiness generally increases with randomness, but not without limit. Continuity is less affected, but drops if matrices become too random. Sparseness has little effect on continuity, but decreases trustworthiness. Decrease in quality appears sublinear, which suggests that sparse t-SNE could be made subquadratic in complexity without too much effect on quality.

Original languageEnglish
Title of host publicationProceedings of the 18th European Symposium on Artificial Neural Networks - Computational Intelligence and Machine Learning, ESANN 2010
Number of pages6
Publication statusPublished - 1 Dec 2010
MoE publication typeA4 Article in a conference publication

Fingerprint Dive into the research topics of 'Effects of sparseness and randomness of pairwise distance matrix on t-SNE results'. Together they form a unique fingerprint.

Cite this