Utility-driven assessment of anonymized data via clustering

Area of Science:

• Data Science
• Social Sciences

Background:

• Clustering is valuable for identifying groups of interest in datasets.
• Data anonymization is crucial for protecting privacy but can impact data utility.
• Assessing the utility of anonymized data for analytical tasks like clustering is essential.

Purpose of the Study:

• To explore clustering techniques as data utility models within data anonymization frameworks.
• To evaluate the impact of anonymization (k-anonymity, differential privacy) on clustering results.
• To assess the utility of anonymized data using metrics, group characteristics, and relative risk.

Main Methods:

• Applied partitional clustering to a dataset of Portuguese higher education Law students.
• Compared anonymized clustering scenarios against the original data.
• Utilized clustering validity indices and standard metrics to evaluate data utility and structure preservation.
• Examined relative risk as a metric for social sciences research.

Main Results:

• For low dimensionality/cardinality datasets, anonymization procedures significantly jeopardize clustering.
• Evidence suggests that anonymized data leads to biased estimates of field-of-study.
• The effectiveness of clustering as a data utility model in anonymization is challenged by data characteristics.

Conclusions:

• Data anonymization can severely impact the integrity of clustering analyses, particularly for smaller datasets.
• The utility of anonymized data for inferring group characteristics, such as field of study, is questionable.
• Careful consideration of anonymization methods and dataset properties is necessary to maintain analytical validity.