Classical ML, validation, metrics, interpretation, debugging · insight-1180
Cluster in the original feature space, not on a t-SNE map, and use the map only to look
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A t-SNE picture of customer or booking features shows islands, and k-means is run on its two coordinates or segment sizes are read off the plot.
Action
Cluster on the scaled original features or a PCA reduction, then colour the t-SNE map with those labels as a visual check. Rerun with several perplexities and seeds and trust only structure that appears in all of them.
Reason
The embedding preserves local neighbourhoods, not global geometry, so downstream numbers computed on it inherit the distortion.
Boundary
Exploratory visualization of embeddings and engineered feature matrices.
Symptom or risk
t-SNE adapts to local density, so cluster size and between-cluster distance carry no meaning, and noise can form clumps at some perplexities; segments change with settings.
Exceptions
When the 2-D map itself is the product, document its settings and treat it as an illustration.
Trade-offs
Several runs cost time and a stable story is less dramatic than one pretty plot.
Pitfalls
Quoting segment shares from the picture rather than from the cluster assignment table.
Lessons and concepts
Claims and their sources
A Distill article demonstrates that in t-SNE plots cluster sizes mean nothing, distances between clusters may not mean anything, perplexity changes the picture, and random noise can look like structure.