COMPARISON OF THE STABILITY OF PROBABILISTIC ID MODELS IN CONDITIONS OF STRICT COMPLIANCE WITH CONFIDENTIALITY REQUIREMENTS

The article examines the stability of probabilistic ID models used in AdTech and programmatic procurement for probabilistic identification and user signal binding (identity resolution) in the context of increased confidentiality requirements. It is proved that privacy constraints (feature minimization, de-identification, aggregation and addition of randomness, including mechanisms of differential privacy) change the information environment of the model and can reduce the stability of probabilistic estimates and threshold solutions with fluctuations in data quality and shifting distributions. A stability comparison framework is proposed that includes a reproducible protocol for preparing and comparing records, modeling typical data degradation (increased gaps, noise, sparsity), and statistical comparison of models using integral quality metrics and calibration indicators with uncertainty estimation. The practical significance of the work lies in the formation of an approach to the selection and configuration of probabilistic ID models that ensure reproducibility of marketing estimates (coverage/frequency, deduplication, attribution, performance measurement) with limited access to identifiers.