Interpretable Parametric Neighbor Embedding

Edouard Couplet, Pierre Lambert, John Aldo Lee, Cyril de Bodt

Track: Type E (Late-Breaking Abstracts)

Keywords: Dimensionality Reduction, Neighbor Embedding, Interpretability

Abstract: Neighbor embedding methods effectively preserve local structures in low-dimensional spaces but are difficult to interpret due to their nonlinear nature. Post-hoc explanations offer only approximate insights. We propose an interpretable neighbor embedding approach that projects each point via a linear combination of shared basis directions, yielding exact explanations through global bases and point-specific coefficients. We demonstrate the method using a t-SNE loss on a single-cell dataset.

Submission Number: 94