Year: 2017
Type: dissertation
Abstract: In this dissertation, I am going to introduce my work on differentially privatedata mining. There are usually two types of privacy preserving data mining algorithms:query answering algorithms and data publishing algorithms.We will present three algorithms on two differentially private query answeringproblems in Chapter 3. The first problem is to return SNPs most correlated to a disease,and we provide two efficient algorithms which make an accurate but previously inefficientalgorithm feasible. The second problem is to learn a model minimizing empiricalrisk while selecting features. Our algorithm beats start-of-the-art algorithms.Then we present three algorithms on differentially private data publishing undertwo scenarios in Chapter 4. The first algorithm assumes existence of public data, andassigns weights to the public data so that they are statistically similar to the privatedata. Analysis on the weighted public data is more accurate than doing analysis onthe private data directly. The two following algorithms assume that published data arefor supervised learning (one algorithm for classification and the other for regression),and that the prediction rules are continuous with respect to predictors. The data thesealgorithms published perform very well on learning tasks.
Author: Zhanglong Ji
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2761345629
