Given a dataset ∈ ℝ^N × D with N samples and D ≫ 10⁴, we seek a mapping 𝑓: ℝ^D → ℝ^d (d ≪ D) such that for any pair (i, j):
All baselines were run on the same 64‑GPU cluster (NVIDIA A100) with equivalent memory budgets. xfredhd
To preserve local geometry, we construct a G on the sketch space using approximate nearest neighbors (ANN) with locality‑sensitive hashing (LSH). The regularizer penalizes deviation of edge similarity after embedding: Given a dataset ∈ ℝ^N × D with