A Multi-objective Optimisation Design Approach for Ensemble Member Combination in Binary Classification Tasks

Machine learning algorithms are usually trained by means of optimising a single global score, and it is possible to have a poor performance when dealing with classification problems where a lower type I or type II error rate is more desirable. To deal with this situation, multi-objective optimisation techniques can be applied for finding the trade-off among the different objectives. However, such approach results in a set of non-dominated solutions, and it is important to select one preferable solution. Thus, this paper proposes the application of a holistic multi-objective optimisation design procedure for ensemble member combination. To do so, a multi-objective problem is formulated in order to maximise the class-specific recalls of a weighted majority voting ensemble. Such problem is optimised with a multi-objective evolutionary algorithm. Next, a ranking technique is used find a preferred solution based on the during the multi-criteria decision making step.