Guided Stable Dynamic Projections

Abstract

Projections aim to convey the relationships and similarity of high-dimensional data in a low-dimensional representation. Most such techniques are designed for static data. When used for time-dependent data, they usually fail to create a stable and suitable low dimensional representation. We propose two dynamic projection methods (PCD-tSNE and LD-tSNE) that use global guides to steer projection points. This avoids unstable movement that does not encode data dynamics while keeping t-SNE's neighborhood preservation ability. PCD-tSNE scores a good balance between stability, neighborhood preservation, and distance preservation, while LD-tSNE allows creating stable and customizable projections. We compare our methods to 11 other techniques using quality metrics and datasets provided by a recent benchmark for dynamic projections.