Year: 2025
Type: dissertation
Abstract: Collecting data is relatively easy, but finding meaningful information from the data requires efficient and thorough data analysis techniques.Pattern mining is a subfield of data analysis, which discovers interesting and interpretable relationships and structures withing the data through various methods.However, these methods come with challenges related to data sensitivity, computational efficiency, and result interpretability.This thesis addresses these challenges in pattern mining with a focus on redescription mining, a pattern mining technique that aims to describe overlapping subsets of data through two disjoint sets of attributes.To improve the accessibility and usability of pattern mining methods, this thesis introduces three approaches which use approximation and randomness to trade off some of the accuracy of the results for gains in privacy and efficiency.The first challenge is pattern mining from sensitive data, such as medical data.Even though the results obtained from a pattern mining algorithm are only patterns and not the original data, they can reveal the private information of the individuals.We solve this problem by combining differential privacy with redescription mining, aiming to provide researchers to access data while protecting the privacy of participants in the data.The proposed algorithm allows for mining good-quality patterns while preserving privacy.The second challenge is scalability of pattern mining methods which rely on nearest neighbor search to high-dimensional datasets.The greedy approach for redescription mining becomes very slow when it has to evaluate several low-quality intermediate candidates.We solve this problem by combining locality-sensitive hashing with greedy redescription mining.The algorithm proposed is able to achieve a significant improvement in the running time, while producing accurate results.The third challenge is choosing an informative and interpretable set of patterns from a very large collection of patterns.Pattern mining methods often result in a very large and overlapping set of patterns.This limits the interpretability of the results, as it is very difficult to identify which patterns are the most informative and interesting.Existing methods for pattern set selection provide accurate results, but can be computationally expensive.We propose an algorithm for fast pattern set selection, and show that we are able to select very fast a set of pattern that give a small reconstruction error.
Author: Karjalainen, Maiju
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7215602546
