Reliable and interactive neighbour embedding for dimensionality reduction, visualisation, and machine learning

Neighbour embedding techniques help analysts make sense of high-dimensional data by representing them in 2- or 3-dimensional spaces, allowing the visualisation of their underlying structures. This work tackles some of their limitations by accelerating the algorithms and adapting them for continual learning, enabling interactive exploration of large datasets. The thesis also investigates why these methods are so effective at separating complex patterns in high-dimensional spaces, arguing that machine learning can gain valuable lessons from neighbour embeddings. This claim is supported by a new label‑propagation algorithm that outperforms comparable methods.