[Paper Review] A Survey on Data Cleaning Methods for Improved Machine Learning Model Performance

Ga Young Lee, Lubna Alzamil+2 more|arXiv (Cornell University)|Sep 15, 2021

Data Quality and ManagementDecision Sciences5 references29 citations

TL;DR

A survey of state-of-the-art data cleaning approaches for improving ML performance, discussing methods like SampleClean, ActiveClean, Holoclean, AlphaClean, and CPClean, their strengths/limitations, and directions for future research.

ABSTRACT

Data cleaning is the initial stage of any machine learning project and is one of the most critical processes in data analysis. It is a critical step in ensuring that the dataset is devoid of incorrect or erroneous data. It can be done manually with data wrangling tools, or it can be completed automatically with a computer program. Data cleaning entails a slew of procedures that, once done, make the data ready for analysis. Given its significance in numerous fields, there is a growing interest in the development of efficient and effective data cleaning frameworks. In this survey, some of the most recent advancements of data cleaning approaches are examined for their effectiveness and the future research directions are suggested to close the gap in each of the methods.

Motivation & Objective

• Motivate the importance of high-quality data for ML performance and address the practical challenges of data cleaning.
• Review recent data cleaning approaches and compare their strengths, weaknesses, and applicability to ML tasks.
• Identify open problems and future research directions in scalable, efficient, and generalizable data cleaning.

Proposed method

• Review and synthesize literature on data cleaning frameworks within data management systems.
• Describe representative approaches (SampleClean, ActiveClean, Holoclean, AlphaClean, CPClean) and their core mechanisms.
• Highlight tradeoffs between coverage and efficiency, and discuss optimizer limitations and generalizability.
• Summarize open problems and propose future research directions across visualization, programming integration, and hardware considerations.

Experimental results

Research questions

• RQ1 What are the dominant data cleaning approaches proposed since 2015 for improving ML model performance?
• RQ2 What are the key strengths and limitations of prominent data cleaning frameworks (SampleClean, ActiveClean, Holoclean, AlphaClean, CPClean) in ML contexts?
• RQ3 What are the major open problems and future directions in data cleaning for scalable and effective ML pipelines?

Key findings

• Data cleaning is costly but essential to reliable ML performance, with dirty data causing substantial inefficiencies and potential revenue impact.
• Recent frameworks aim to reduce human effort and improve scalability, using approaches like simulated clean data, incremental learning, probabilistic inference, and pipeline generation.
• Tradeoffs exist between full data coverage and computational efficiency, influencing method selection based on dataset characteristics.
• Optimizer design and user interaction are critical barriers for practical adoption and generalization across domains.
• Future directions emphasize visualization, unified probabilistic data programming, and hardware-enabled memory management to enhance performance and usability.