Efficient $$F$$F measure maximization via weighted maximum likelihood

摘要: The classification models obtained via maximum likelihood-based training do not necessarily reach the optimal $$F_\beta $$Fs-measure for some user's choice of $$\beta $$s that is achievable with chosen parametrization. In this work we link weighted entropy and optimization expected $$Fs-measure, by viewing them in framework a general common multi-criteria problem. As result, each solution maximization can be realized as likelihood within model - well understood behaved problem which standard (off shelf) gradient methods used. Based on insight, present an efficient algorithm $$Fs using dynamically adaptive weights.