Human Preference aligned Tabular Similarity

Short summary

This paper argues that standard downstream metrics are insufficient for evaluating tabular embeddings used in similarity search, as they don't capture human preference alignment. The authors propose a concrete evaluation procedure for human-preference-aligned similarity rankings and demonstrate the problem through a Product Lifecycle Management use case. The contribution is primarily an evaluation methodology rather than a new embedding model.

• •Standard metrics insufficient for assessing tabular embedding trustworthiness in similarity search
• •Proposes human-preference-aligned evaluation procedure
• •Illustrates problem through a PLM use case

Generated with AI, which can make mistakes.