@online{heiter:21:factoring,
TITLE = {Factoring Out Prior Knowledge from Low-dimensional Embeddings},
AUTHOR = {Heiter, Edith and Fischer, Jonas and Vreeken, Jilles},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2103.01828},
EPRINT = {2103.01828},
EPRINTTYPE = {arXiv},
YEAR = {2021},
ABSTRACT = {Low-dimensional embedding techniques such as tSNE and UMAP allow visualizing<br>high-dimensional data and therewith facilitate the discovery of interesting<br>structure. Although they are widely used, they visualize data as is, rather<br>than in light of the background knowledge we have about the data. What we<br>already know, however, strongly determines what is novel and hence interesting.<br>In this paper we propose two methods for factoring out prior knowledge in the<br>form of distance matrices from low-dimensional embeddings. To factor out prior<br>knowledge from tSNE embeddings, we propose JEDI that adapts the tSNE objective<br>in a principled way using Jensen-Shannon divergence. To factor out prior<br>knowledge from any downstream embedding approach, we propose CONFETTI, in which<br>we directly operate on the input distance matrices. Extensive experiments on<br>both synthetic and real world data show that both methods work well, providing<br>embeddings that exhibit meaningful structure that would otherwise remain<br>hidden.<br>},
}