Output details
11 - Computer Science and Informatics
University of Manchester
Beyond Fano's Inequality: Bounds on the Optimal F-Score, BER, and Cost-Sensitive Risk and Their Implications
<13> Maximum likelihood is widely applied in statistics and machine learning, commonly as a proxy to maximise the accuracy of a predictor. This paper is significant as it proves for the first time that similar proxies can be derived for more complex measures such as F-score or cost-sensitive risk. It is shown that maximum likelihood, though widely applied in machine learning to maximise F-score, can lead to catastrophic failures; we suggest an alternative proxy measure that overcomes this. This has implications for any work that has attempted to use maximum likelihood techniques with the intention of optimizing F-score.