ENH: Support vector-valued metadata and relevant filtering operations

[tcoffman-strive]

Feature Type

• Adding new functionality to valor

• Changing existing functionality in valor

• Removing existing functionality in valor

Problem Description

I wish I could leverage semantic monitoring and/or other rich decompositions of my input space, together with Valor slicing and filtering capabilities, to gain deeper understanding into model performance in a more automated way.

Feature Description

Current capabilities allow user-defined metadata. In theory, vector-valued metadata can be added by just adding N scalar-valued metadata fields. But this significantly limits the depth and relevance of analyzing those values. If I could instead submit vector-valued metadata, and then Valor further supported relevant filtering, distance metrics (Euclidean, p-norms, Mahalanobis distance), or analysis techniques (clustering, multi-dimensional scaling, density estimation, t-SNE, UMAP, etc.), I could file semantic interpretations (e.g., embeddings) as metadata, and likely uncover more subtle and meaningful performance trends (e.g., "This model performs poorly when there is snow on the ground", etc.).

Additional Context

No response