Clustering Heterogeneous Data Values for Data Quality Analysis

Data is of high quality if it is fit for its intended purpose. Data heterogeneity can be a major quality problem, as quality aspects such as understandability and consistency can be compromised. Heterogeneity of data values is particularly common when data is manually entered by different people using inadequate control rules. In this case, syntactic and semantic heterogeneity often go hand in hand. Heterogeneity of data values may be a direct result of problems in the acquisition process, quality problems of the underlying data model, or possibly erroneous data transformations. For example, in the cultural heritage domain, it is common to analyze data fields by manually searching lists of data values sorted alphabetically or by number of occurrences. Additionally, search functions such as regular expression matching are used to detect specific patterns. However, this requires a priori knowledge and technical skills that domain experts often do not have. Since such datasets often contain thousands of values, the entire process is very time-consuming. Outliers or subtle differences between values that may be critical to data quality can be easily overlooked. To improve this process of analyzing the quality of data values, we propose a bottom-up human-in-the-loop approach that clusters values of a data field according to syntactic similarity. The clustering is intended to help domain experts explore the heterogeneity of values in a data field and can be configured by domain experts according to their domain knowledge. The overview of the syntactic diversity of the data values gives an impression of the rules and practices of data acquisition as well as their violations. From this, experts can infer potential quality issues with the data acquisition process and system, as well as the data model and data transformations. We outline a proof-of-concept implementation of the approach. Our evaluation found that clustering adds value to data quality analysis, especially for detecting quality problems in data models.