Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2026
- Goals
- Record
- 0087e2b3-7e48-491a-b911-dbe6973d97e3
- Published
- 2026-10-30
Universitas Dian Nuswantoro conducted research on rare-event classification in noisy, highly imbalanced tabular data, with Minardi, Joko; Budiman, Fikri; Fanani, Ahmad Zainul; and Rijati, Nova participating as researchers. The research addressed routine growth-monitoring records used for scalable screening of severe child undernutrition. It identified that reliable learning is difficult because severe cases are rare and routine measurements introduce feature and label noise.
The activity proposed TAMS-DCSL, a two-stage framework. Topology-Aware Minority Sampling generated synthetic minority instances while preserving local neighborhood structure. Sampling was restricted to safe-core minority regions identified through neighborhood density and label consistency, reducing noise and class overlap. Dynamic Cost-Sensitive Learning updated misclassification costs during training using validation feedback, addressing the brittleness of static cost ratios.
UDINUS evaluated the framework on a child growth-monitoring dataset with 47,964 records and a composite severe anthropometric failure label based on WHO severe z-score criteria. Stratified k-fold results improved minority recall from 0.643 to 0.785 and GMean from 0.801 to 0.863 compared with standard XGBoost while maintaining high specificity. Public benchmarks confirmed robustness under extreme imbalance and noisy label conditions. The activity therefore presents an institutional research effort focused on improving detection of severe anthropometric failure in routine monitoring data.