MEG: Multi-objective Ensemble Generation

Recent studies have found that ensemble prediction models (i.e., aggregation of multiple base classifiers) can achieve more accurate results than those that would have been obtained by relying on a single classifier when used to make predictions. However, designing an ensemble requires a non-trivial amount of effort and expertise with respect to the choice of the set of base classifiers, their hyper-parameter tuning, and the choice of the strategy used to aggregate the predictions. An inappropriate choice of any of these aspects can lead to over- or under-fitting, thereby heavily worsening the performance of the prediction model. Examining all possible combinations is not computationally affordable, as the search space is too large, and there is a strong interaction among these aspects, which cannot be optimised separately.

In this talk, I will present a novel approach based on multi-objective evolutionary search to automatically generate optimal ensemble classifiers. We dubbed this approach as Multi-objective Ensemble Generation (MEG). We verify the effectiveness of MEG in detecting defects across traditional software versions as well as in detecting a new type of defect (i.e., bias) in machine learning software systems. This talk will cover MEG as well as its application in Software Defect Prediction and Software Fairness.