TriMap is a dimensionality reduction method that forms a low-dimensional embedding of data by minimizing a contrastive loss over a set of triplets. The triplets are sampled from the original high-dimensional data representation and are weighted based on the distances between the (closer and farther) pairs of points. Although t-SNE and UMAP are excellent methods for forming low-dimensional embeddings, TriMap provides an alternative view of the data which is more representative “globally”. Specifically, TriMap is able to:

1. reflect the relative placement of the clusters in high-dimension,
2. reveal possible outliers and anomalies in the data,
3. generate embeddings that are more robust to certain transformations (see here for more details).