Résumé:
With advancements in data analysis and processing techniques, the release of micro-data for
research purposes, such as disease outbreak studies or economic pattern analysis, has become
prevalent. However, these datasets, while valuable for researchers, often contain sensitive information
that poses privacy risks to individuals. Privacy-preserving data mining (PPDM) has
emerged as a critical field to address these concerns by concealing sensitive information while
still enabling the extraction of useful insights. This task is NP-hard and involves the challenge
of balancing the concealment of sensitive itemsets with the preservation of non-sensitive ones
during data extraction. Numerous algorithms have been developed for deterministic databases,
where information is binary (present or absent). This thesis explores a novel PPDM approach
in the context of uncertain databases, where information is represented by probabilistic values.
The sanitization process, aimed at hiding sensitive information, introduces side effects such as
hiding failure, missing cost, artificial cost, and dissimilarity. These side effects are considered
as objective functions to be minimized. To achieve the goal of PPDM, a Multi-Objective Optimization
problem is formulated, and metaheuristic algorithms are employed. Specifically, the
NSGA-II algorithm is applied, leading to the development of the NSGAII4ID algorithm for
deterministic databases. This algorithm hides sensitive frequent itemsets by removing selected
items from chosen transactions, representing a pioneering effort in privacy-preserving data mining.
Furthermore, for uncertain databases, a novel algorithm named U-NSGAII4ID is proposed,
addressing the multi-objective optimization problem by encoding a set of items to remove
from selected transactions. Additionally, three heuristic approaches for PPDM in uncertain
databases are introduced: the aggregate approach, which removes transactions; the disaggregate
approach, which removes selected items from each transaction; and the hybrid approach,
combining the two former approaches. Experimental evaluations compare these approaches,
demonstrating their effectiveness in preserving sensitive itemsets in uncertain databases. *
