PCA and t-SNE for Exploring MailTrace’s Embedding Space

Ask what the projection preserves

PCA finds linear directions of variation. t-SNE emphasises local neighbourhood structure and can produce different arrangements under different settings. Neither guarantees that visible islands correspond to distinct requests, legal categories or approved issues.

The useful application is diagnostic. Colour points by an independently reviewed family, inspect known false merges or identify repeated boilerplate. If a large island consists of signature text, the problem may be preparation rather than the clustering algorithm. A projection can suggest where to investigate without deciding the case.

Keep inspection linked to original source IDs

This sketch shows the planned steps rather than a runnable data pipeline. Actual experiments need enough examples, a stated embedding generation and parameters appropriate to the sample size. Sensitive evidence should stay in an authorised environment.

1. Select one compatible embedding generation.
2. Retain a source ID for every vector.
3. Fit the chosen projection for exploration.
4. Colour by reviewer labels, not model predictions alone.
5. Inspect selected points in their original evidence context.
6. Compare findings in the original vector space.

Do not tune a screenshot into a benchmark

Trying several settings and choosing the most attractive plot can conceal instability. Record the algorithm, seed and parameters, and compare neighbourhood behaviour across reasonable choices. Avoid interpreting distances between distant t-SNE islands as a calibrated measure of difference.

If projection outputs become features for a classifier, fit the transformation only within the training procedure and evaluate its effect separately. For ordinary product review, retain the original representations for retrieval and constraints. The picture should help people ask better questions about the evidence pipeline, not become another unreviewed source of truth.