Can one dimension keep the neighborhoods?
→ ← same neighborhood attracts ← → different neighborhoods repel
step 0
Interactive two-dimensional to one-dimensional neighborhood embedding Three clusters are separate in two dimensions but overlap after direct projection to one dimension. Repeated updates attract members of the same neighborhood and repel members of different neighborhoods. Original space: 2D feature 1 feature 2 embed Target space: 1D projection direction one learned coordinate
Raw projection: colors overlap even though the 2D neighborhoods do not.
Teaching simplification: color stands in for similarity in the original space. Real t-SNE uses pairwise similarities—not cluster labels.