⚖️ Cost-sensitive Classification
Overview
Not all mistakes are equal!
In FungiTastic, confusing a poisonous species for an edible one can have severe consequences, whereas the reverse is less dangerous.
This benchmark evaluates models under custom loss/cost functions that reflect real-world priorities.
Use Cases
- Safety-critical systems (e.g., mobile ID for foragers)
- Research into cost-sensitive, risk-aware, or calibrated classifiers
Data & Splits
- Uses standard FungiTastic splits (temporal)
- Poisonous/edible labels and additional risk metadata provided
Evaluation Protocol
- Metrics: Weighted error rates, custom cost matrices (see Appendix in the paper)
- You may define your own costs for specific errors
Baselines & Results
- Example: False positive edible has high penalty; false positive poisonous has lower penalty
- See example code and detailed description in Baselines & Models and evaluation protocols
Quick Start
- Use provided cost matrix and baseline scripts
- Extend to new tasks with your own risk functions