Semantic map

Every review, on one canvas.

47,842 reviews across every indexed app, projected to 2D by UMAP. Each dot is one review.

How to read this map

Every review's 1024-dimensional embedding has been squeezed into 2D by UMAP, a non-linear dimensionality reduction algorithm. Reviews land near each other when the model thinks they're talking about the same thing — so visible clusters are real semantic groups, not arbitrary placements. The distance between two clusters roughly tracks how related the underlying complaints/praise are; the shape of a cluster tells you whether the conversation is tight (one specific gripe) or sprawling (a category of related ones). Color is the star rating: red 1★, amber 2-3★, green 4-5★. Drag to pan, scroll to zoom, hover for the review text.

Loading 47,842 reviews…
1★ 2-3★ 4-5★
zoom: 1.00x