Optimal threshold selection algorithms for multi-label classification: property study

Multi-label classification models arise in various areas of life, which is explained by an increasing amount of information that requires prompt analysis. One of the mathematical methods for solving this problem is a plug-in approach, at the first stage of which, for each class, a certain ranking function is built, ordering all objects in some way, and at the second stage, the optimal thresholds are selected, the objects on one side of which are assigned to the current class, and on the other — to the other. Thresholds are chosen to maximize the target quality measure. The algorithms which properties are investigated in this article are devoted to the second stage of the plug-in approach which is the choice of the optimal threshold vector. This step becomes non-trivial if the $F$-measure of average precision and recall is used as the target quality assessment since it does not allow independent threshold optimization in each class. In problems of extreme multi-label classification, the number of classes can reach hundreds of thousands, so the original optimization problem is reduced to the problem of searching a fixed point of a specially introduced transformation V, defined on a unit square on the plane of average precision $P$ and recall $R$. Using this transformation, two algorithms are proposed for optimization: the $F$-measure linearization method and the method of V domain analysis. The properties of algorithms are studied when applied to multi-label classification data sets of various sizes and origin, in particular, the dependence of the error on the number of classes, on the $F$-measure parameter, and on the internal parameters of methods under study. The peculiarity of both algorithms work when used for problems with the domain of V, containing large linear boundaries, was found. In case when the optimal point is located in the vicinity of these boundaries, the errors of both methods do not decrease with an increase in the number of classes. In this case, the linearization method quite accurately determines the argument of the optimal point, while the method of V domain analysis — the polar radius.