Revised Conditional t-SNE: Looking Beyond the Nearest Neighbors

Conditional t-SNE (ct-SNE) is a recent extension to that allows removal of known cluster information from the embedding, obtain visualization revealing structure beyond label information. This useful, for example, when one wants factor out unwanted differences between set classes. We show ct-SNE fails in many realistic settings, namely if data well clustered over labels original high-dimensional space. introduce revised method by conditioning similarities instead low-dimensional and storing within- across-label nearest neighbors separately. also enables use recently proposed speedups t-SNE, improving scalability. From experiments on synthetic data, we find our resolves considered problems improves embedding quality. On real containing batch effects, expected improvement not always there. argue preferable overall, given its improved The results highlight new open questions, such as how handle distance variations clusters.