Naslov
On Differentiating Synthetic and Real Data in Medical Applications

Autori
Baressi Šegota, Sandi ; Anđelić, Nikola ; Štifanić, Daniel ; Štifanić, Jelena ; Car, Zlatan

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
The Second Serbian International Conference on Applied Artificial Intelligence (SICAAI) - Book of Abstracts / Filipović, Nenad - Kragujevac : University of Kragujevac, 2023

ISBN
978-86-81037-77-5

Skup
The Second Serbian International Conference on Applied Artificial Intelligence (SICAAI)

Mjesto i datum
Kragujevac, Srbija, 19.05.2023. - 20.05.2023

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
data classification, machine learning, synthetic data.

Sažetak
The use of synthetic, generated, data to address the machine-learning algorithms’ needs for a large amount of data points is a growing trend in the research community. This paper tests four methods - Copula Generative Adversarial Network (GAN), CTGAN, Gaussian Copula, and Triplet-based variable encoder (TVAE) on the dataset defining risk factors for lung cancer - selected due to the high performance of ML models applied to it. The resulting datasets are tested using Pearson’s correlation coefficient and used for training two multilayer perceptron (MLP)-based neural networks. The first tests the classification performance of the models developed with synthetic datasets compared to the ones developed on original data. The second network is trained with the goal of differentiating between real and synthetic data through training on a mixed dataset. Results show that the correlation similarity has a direct link with the performance of both networks, with TVAE generating the best results and being the hardest to differentiate.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo, Interdisciplinarne tehničke znanosti

POVEZANOST RADA