Autoencoder based generators of semi-artificial data

Klančar, Jaka (author), Robnik Šikonja, Marko (mentor)

The goal of the thesis is to alleviate the problem of insufficient data available for data analysis or machine learning. We developed a generator of semi-artificial data based on autoencoders. We implemented dynamic autoencoders without any predefined structure, as we wanted that our solution is general and may therefore be used on any data set. Results showed that autoencoder based generators work better than variational autoencoders. The generators perform best on data sets with a small number of mixed attributes and balanced classes. They perform better if more training instances are available. Results additionally show that grid search significantly improves the performance and that it is possible to predict a good set of parameters for each data set.

• Language: English
• Keywords: autoencoders, data generators, neural networks, semi-artifical data, variational autoencoders
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-102239
• Publication date in RUL: 26.07.2018

Secondary language

• Language: Slovenian
• Title: Generatorji delno umetnih podatkov na podlagi samokodirnikov

Abstract

Glavni cilj naloge je bil olajšati problem pomanjkanja podatkov pri analizi podatkov in v strojnem učenju. Razvili smo generator delno umetnih podatkov na podlagi samokodirnikov. Implementirali smo dinamične samokodirnike brez vnaprej določene strukture, saj smo želeli, da so generatorji uporabni na poljubni učni množici. Rezultati so pokazali, da generatorji na podlagi samokodirnikov delujejo bolje kot variacijski samokodirniki. Naši generatorji najbolje delujejo na podatkovnih množicah z manjšim številom atributov in z uravnoteženimi razredi. Večje število učnih primerov izboljša delovanje generatorjev. Rezultati so tudi pokazali, da z mrežnim iskanjem znatno izboljšamo rezultate in da je možno napovedati dobre parametre glede na karakteristike dane podatkovne množice.

• Keywords: samokodirniki, generatorji podatkov, nevronske mreže, delno umetni podatki, variacijski samokodirniki