Knowledge-Guided Data Cleaning: From Knowledge Acquisition to Reliable Decisions

Data cleaning requires identifying errors and determining which changes the available evidence supports. Integrity constraints reveal inconsistencies, while trusted reference data, statistical regularities, and human feedback help locate errors and select corrections. This survey examines how knowledge supports these decisions, connecting its representation, acquisition, and reuse with candidate generation, evidence combination, and repair selection. We compare methods according to how they obtain knowledge, establish its applicability, and respond when evidence conflicts or confidence is insufficient. The comparison distinguishes factual correctness, constraint consistency, and downstream task utility, which lead systems to select different operations and require different evaluation criteria. It also explains how experimental assumptions about known error locations, candidate coverage, and accepted predictions affect comparisons across studies. As machine learning and large language models extend cleaning into table transformation and preparation planning, knowledge must support both local repairs and coordination across processing steps. Building on these connections, we identify research directions in checking knowledge applicability, selecting responses to conflicts, and acquiring information according to unresolved decisions.