Learning from Anonymized and Incomplete Tabular Data

TL;DR

The paper addresses learning from tabular data where user driven privacy yields mixtures of original, generalized, and missing values. It formalizes the problem with dataset D and anonymized D^A, introducing a data preparation function phi and an objective that minimizes the utility gap between learning on original data and anonymized data, phi* = argmin_phi in Phi |U(f^*) − U(f_phi)|. The authors propose granularity based imputation via specialization, along with baselines such as no prep and standard imputation, and compare against LLM based methods across four diverse datasets and deployment scenarios AnTr, AnTe, and AnBo, under various privacy distributions. Key results show that generalized values generally preserve more utility than suppression, specialization is the most robust prep across scenarios, and method choice is highly scenario dependent; in some extreme cases, LLM based approaches can compensate for the absence of original values. The work demonstrates that effective learning under user driven privacy can be achieved through careful data preparation on anonymized data, enabling privacy aware ML without altering models or reversing anonymization, and points to future work on real-world data and improved privacy metrics for heterogeneous anonymization.

Abstract

User-driven privacy allows individuals to control whether and at what granularity their data is shared, leading to datasets that mix original, generalized, and missing values within the same records and attributes. While such representations are intuitive for privacy, they pose challenges for machine learning, which typically treats non-original values as new categories or as missing, thereby discarding generalization semantics. For learning from such tabular data, we propose novel data transformation strategies that account for heterogeneous anonymization and evaluate them alongside standard imputation and LLM-based approaches. We employ multiple datasets, privacy configurations, and deployment scenarios, demonstrating that our method reliably regains utility. Our results show that generalized values are preferable to pure suppression, that the best data preparation strategy depends on the scenario, and that consistent data representations are crucial for maintaining downstream utility. Overall, our findings highlight that effective learning is tied to the appropriate handling of anonymized values.

Figures (6)

• Figure 1: Overview of evaluated preparation methods for anonymized datasets with generalized and missing values.
• Figure 2: Comparison of possible preparation outcomes on an anonymized data sample, where $\bot$ denotes missing values.
• Figure 3: F1 scores for preparing methods across evaluation scenarios, datasets, and privacy distributions. The x-axis shows the distributions as Original-Generalized-Missing percentages. Dashed lines indicate baseline performance on original data.
• Figure 4: Heatmap of average F1 utility loss against original baseline for the three best preparation methods across privacy distributions excluding 0-66-34 as a special case.
• Figure 5: Heatmap of average F1 utility loss against original baseline for the three best preparation methods for the 0-66-34 distribution only (no original values).