Wonjae Lee

Wonjae Lee

Auto-generated profile, based on publicly available information

Wonjae Lee


Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

12 May 2025 in Sustainable Energy, Grids and Networks
Verified
Journal Article
Closed access
Kangwon SeoKangwon Seo
Kangwon Seo
HNHyeong Suk Na
Hyeong Suk Na
WLWonjae Lee
Wonjae Lee
CCCHENG-BANG CHEN
CHENG-BANG CHEN
SKSang Jin Kweon
Sang Jin Kweon
SKSoundar Kumara
Soundar Kumara

14 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
KLKyung-Sun Lim
Kyung-Sun Lim
WLWonjae Lee
Wonjae Lee
JCJae-Hyuk Choi
Jae-Hyuk Choi
Every acquisition function in common use rests on an assumption class-conditional generation violates: that a selected sample's label is unknown and must be bought from an oracle. That assumption is what makes uncertainty the quantity to maximise. When candidates are drawn from a generator that was told which class to synthesise, the label is free, and the criterion inverts --- an uncertain candidate is no longer information awaiting purchase but a sample whose assigned class is probably wrong, that is, label noise. This paper takes that inversion as its premise and develops oracle-free acquisition for generative augmentation under severe class imbalance: the candidates worth keeping are those a classifier confidently assigns to the class they were conditioned on, and are mutually diverse. The rule puts the classifier inside the generative loop rather than after it, realised in four coupled stages --- an autoencoder that initialises the generator and supplies a class-conditional latent prior, a gradient-penalised adversarial backbone, interpolation-seeded candidates filtered by the oracle-free rule, and co-training of critic, generator and classifier anchored to a frozen reference that stops the generator chasing an imperfect classifier. An ablation attributes performance to each stage. Against five established generative augmenters on three public image datasets under a shared protocol, the method ranks first on all four metrics on two datasets and leads six of the nine cells of an imbalance-ratio sweep. The results further show that realism and utility are separate objectives: the augmenter with the best Fr\'echet Inception Distance returns the worst macro-$F2$.