“Visualizing Data using t-SNE” by Laurens van der Maaten and Geoffrey Hinton appeared in the Journal of Machine Learning Research, volume 9, in 2008 (pages 2579-2605). It introduced t-distributed stochastic neighbor embedding, or t-SNE, a method for placing high-dimensional data points onto a two- or three-dimensional map for visualization.
t-SNE works by converting distances between points into probabilities: nearby points get high probability of being neighbors, distant points low probability. It then arranges the low-dimensional map so that its neighbor probabilities match the original ones as closely as possible. Using a heavy-tailed Student t-distribution in the map space, the paper’s key refinement over earlier stochastic neighbor embedding, prevents points from crowding into the center and produces clearly separated clusters.
The technique became enormously popular for inspecting what neural networks and other models have learned. Plotting the internal representations of, say, handwritten digits or word embeddings with t-SNE reveals tight, interpretable clusters. Practitioners learned to read its maps with care: t-SNE preserves local neighborhoods well but can distort global distances and cluster sizes.
Why business readers should care: t-SNE is the standard way to turn an opaque, high-dimensional dataset into a picture a human can actually look at and reason about, making it a common first step in exploring customer segments or model behavior.