Agentic Data Cleaning Without Clean Reference: Capabilities and Trade-offs
A recent study published on arXiv (2608.14765) delves into agentic data cleaning in situations where there’s no reliable reference for clean data. This means that unusual values might be mistakes or genuine data points. The researchers propose a new framework that uses evidence and includes various strategies like context structuring, profiling, LLM reasoning, and more. They tested seven different setups on datasets related to finance, healthcare, and environmental monitoring, completing a total of 126 runs. The study compared two baseline methods and utilized a progressive LLM-based approach. Although the deterministic profiling baseline showed promising results, the abstract doesn’t reveal more details. This work addresses a key challenge in data cleaning when a clean reference isn’t present, balancing automation with human input.
Key facts
- Paper arXiv:2608.14765v1, announced as new
- Focuses on data cleaning without a clean reference
- Proposes evidence-grounded framework with 10 components
- Evaluates seven configurations across three dataset types
- Uses controlled synthetic corruption and descriptive analysis
- Includes 126 completed runs
- Two comparison baselines and progressive LLM-based sequence
- Deterministic profiling baseline performed well in synthetic evaluation
Entities
Institutions
- arXiv