mlr3torchAUM A Comprehensive Solution for Imbalanced Classification in R

Problem: Imbalanced datasets pose a significant challenge in machine learning, where standard metrics like accuracy can falsely evaluate a model that simply predicts the dominant class. Currently, the R community lacks a comprehensive and unified framework for handling imbalanced data and X-Risks problems natively integrated into the mlr3/mlr3torch ecosystem. Solution: This project, mlr3torchAUM, provides a comprehensive solution for imbalanced classification in R. It will bridge the current gap by integrating state-of-the-art loss functions, advanced evaluation metrics, and various data sampling techniques directly into the mlr3/mlr3torch pipeline. Deliverables: 1. State-of-the-Art Loss Functions: Implementation of algorithms including log-linear all-pairs squared hinge loss, Area Under Min (AUM) loss, and cost-sensitive logistic loss. 2. Sampling Operators: A robust suite of over-sampling techniques including SMOTE and ADASYN, alongside under-sampling techniques, all implemented safely under a `lazy_tensor` mechanism to avoid memory issues. 3. Advanced Evaluation Metrics: Integration of Precision-Recall curves, Macro/Micro AUM/AUC, and MCP/IMCP to fully support multiclass imbalanced scenarios. 4. LibAUC Integration: Porting cutting-edge losses and optimizers from the Python LibAUC library into R. 5. Engineering & Release: Comprehensive Test Driven Development, GitHub Actions for CI/CD, detailed vignettes, and a final submission to CRAN.