Section: New Results
An Algorithm to Optimize the F-measure by Proper Weighting of Classification Errors
Participants: K. Bascol, R. Emonet, E. Fromont, A. Habrard, G. Metzler, M. Sebban
 proposes an F-Measure optimization algorithm with theoretical guarantees that can be used with any error-weighting learning method. The algorithm, iteratively generates a set of costs from the training set so that the final classifier has an F-measure close to optimal. The optimality of the F-measure is expressed using a finer upper bound as presented in . Furthermore, we show that the costs obtained at each iteration of our method can drastically reduce the search space and thus converge quickly to the optimal parameters. The efficiency of the method is shown both in terms of F-measurement but also in terms of speed of convergence on several unbalanced datasets.