Year: 2021
Type: dissertation
Abstract: Generative models are commonly used to generate realistic data for different purposes such as preserving the privacy of sensitive data or performance and unit testing. The majority of applications of fake data generation concerns itself with images, videos, text, or tabular data. A less explored direction is a particular type of sequential data where the time-stamp or position of elements of the sequence are not at constant intervals, i.e., irregular sequential data. Such data can for example contain the logs of events with time stamps in a company, which may be considered as sensitive information. This thesis proposes new generative models for such irregular sequential data. Variational auto encoder (VAE) models are powerful methods to generate non-sequential fake data. Following the huge success of transformer models in natural language processing applications, the thesis studies the use of attention network and applies it in a VAE model. The new network, called VAE-Attention, shows improved performance and better discriminative scores compared to a baselines and a related work. The analysis in this thesis shows that the model is able to generate high-quality irregular sequential data while having a reasonable training cost. The thesis is concluded with discussions about ways to include attention network in VAE models.
Source: Aaltodoc (Aalto University)
Author: Safar, Shaghayegh
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7028752387