Place: Large Lecture Room
Affiliation: Bangor University, Wales, UK
Combining classifiers is a mature discipline now, which begs the question “Is there any mileage left in it?” This talk will present a bibliographic snapshot of the classifier ensemble area as of January 2013, and will try to answer some aspects of the above question.
We will cast a fresh eye on the diversity-accuracy dilemma via kappa-error diagrams. A limit on the diagram will be presented, which will reveal how much scope for improvement we may still have. Issues such as the importance and rigor of the classifier combiner will be discussed. We will touch upon Bob Duin’s question “to train or not to train” the combiner. Finally, we will indulge in an audience-led little speculation about future research avenues in classifier ensembles.