Non Decomposable Performance Measure

Non-decomposable performance measures, unlike accuracy or cross-entropy, cannot be easily broken down into individual data point contributions, posing challenges for machine learning optimization. Current research focuses on developing algorithms to learn effectively with these measures, particularly in scenarios with noisy labels or fairness constraints, employing techniques like Frank-Wolfe and Bisection methods, as well as coreset sampling for efficient computation. This work is significant because it enables the development of robust and fair classifiers for real-world applications where metrics like F1-score and Matthews Correlation Coefficient are crucial, improving model performance and addressing biases in data.