Dimensionality Reduction Considered Harmful (Some of the time)

Abstract:

Dimensionality reduction (DR) techniques, e.g., t-SNE, UMAP, and PCA, are useful for visually analyzing high-dimensional data. However, DR is often misused in practice, e.g., by cherry-picking hyperparameters, which compromises the reliability of visual analytics and communication. In this talk, we discuss why the misuse occurs and how it can be addressed. We conclude by strongly, yet reluctantly, advocating for automation as a solution to making proper use of DR a norm.

About Hyeon:

Hyeon Jeon (hyeonjeon.com) is a final-year Ph.D. Student at the Department of Computer Science and Engineering, Seoul National University. He aims to make visual analytics and communications more reliable. His research is published in prestigious venues for visualization, machine learning, and human-computer interaction, including TVCG, VIS, TPAMI, and CHI. He received a B.S. in Computer Science and Engineering from POSTECH. His Ph.D. study is supported by the Google Ph.D. fellowship.