Human Preference Aligned Tabular Similarity
A recent study contends that existing tabular embeddings, designed primarily for prediction tasks, do not yield similarity rankings that correspond with human preferences. The researchers assert that conventional downstream metrics do not adequately assess the reliability of embeddings in similarity searches, especially within Product Lifecycle Management (PLM) systems. They introduce a specific evaluation method and demonstrate the issue using a PLM case study. This paper falls under the category of Computer Science > Machine Learning and has been submitted to arXiv.
Key facts
- Tabular embeddings are used for similarity search in PLM.
- Leading embeddings are optimized for prediction, not human-aligned similarity.
- Standard metrics are insufficient for evaluating similarity search trustworthiness.
- A concrete evaluation procedure is presented.
- The problem is illustrated through a PLM use case.
- The paper is on arXiv under Computer Science > Machine Learning.
Entities
Institutions
- arXiv