Resampled Ensemble Learning for Imbalanced Data

Self research project · December 2022 – January 2023

Prototype

Class imbalance 1 : 10 Apply HEM + SEM + BCWF ?

Accuracy
Minority F1

Illustrative numbers — the point the real experiments made: accuracy stays flattering while minority-class F1 collapses, until resampling and example mining rebalance the ensemble.

Highlights

Addressed severe classification data imbalance anomalies by writing customized Python algorithm structures from scratch.
Code-authored Hard Example Mining (HEM) and Soft Example Mining (SEM) algorithms to prioritize and adjust weight criteria of complex, edge-case samples based on training cross-entropy errors.
Built the Balanced Classed Weights and Features (BCWF) balancing script alongside an ensemble voter layer to blend sparse predictive structures, securing significant leaps in overall F1-score and balanced accuracy over basic AdaBoost benchmarks.

Technologies

Python
Hard Example Mining (HEM)
Soft Example Mining (SEM)
BCWF
AdaBoost