Publication Date

Summer 2026

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Applied Data Science

Advisor

Simon Shim; Charles Choo; Guannan Liu

Abstract

Modern machine learning (ML) organizations rely on feature stores to manage training and production data, yet as catalogs grow to thousands of features, semantic management becomes the bottleneck: discovery, governance, and metric selection remain largely manual. This thesis proposes and evaluates a knowledge-enhanced feature store that augments a dual-plane store with a hybrid knowledge layer—relational provenance, a lineage graph, semantic vector retrieval, and an online cache—and an LLM-driven multi-agent layer for profiling, matching, grounded metric recommendation, and pipeline generation. A working prototype was evaluated across three task families—semantic profiling, feature-matching retrieval, and discovery—using classification, ranking, and operational metrics computed by a dynamic suite that derives its criteria from the live registry. Results show that no single representation suffices: vector retrieval supplies recall but cannot enforce constraints, graph structure supplies auditability but is sparse under cold start, and LLM reasoning supplies interpretation but requires grounding. Hybrid retrieval produced better top-ranked results than either method alone. The study concludes that combining vector candidate generation, graph and rule constraints, and grounded LLM explanation offers the best balance for operational ML and LLM workflows.

Download

Share

COinS