A multi-label voting algorithm for neuro-fuzzy classifier ensembles with applications in visual arts data mining

The term visual arts data mining defines a framework for Data Mining techniques applied to learn and discover patterns in visual arts collections. Its results can be widely used by visual arts market, museums and art galleries. This paper proposes a multi-label voting algorithm to identify similar visual arts objects studied using neuro-fuzzy classifiers. The algorithm integrates predictions of experts trained on clusters of heterogeneous collections of data. It combines predictions of the modular ensemble of classifiers by identifying hierarchical votes for most similar classes. Experimental results show better performances than individual global models. Relationships between some visual arts patterns are inferred. We also compare the results obtained for few fusion versions of our algorithm with other methods applied on IRIS and Glass benchmarks. The results show that our algorithm has at least similar performance to other schemes on all data sets and adds flexibility to cases where classifiers' expertise overlaps on unions of disjunctive sets of the universe of discourse.