An Empirical Study of Shape Recognition in Ensemble Learning Context

Shape recognition has been a popular application of machine learning, where each shape is defined as a class for training classifiers that recognize the shapes of new instances. Since training of classifiers is essentially achieved through learning from features, it is crucial to extract and select a set of relevant features that can effectively distinguish one class from other classes. However, different instances could present features which are highly dissimilar, even if these instances belong to the same class. The above difference in feature representation can also result in high diversity among classifiers trained by using different algorithms or data samples. In this paper, we investigate the impact of multi-classifier fusion on shape recognition by using six features extracted from a 2D shape data set. In particular, popular single learning algorithms, such as Decision Trees, Support Vector Machine and K Nearest Neighbours, are adopted to train base classifiers on features selected by using a wrapper approach. Furthermore, two popular ensemble learning algorithms (Random Forests and Gradient Boosted Trees) are adopted to train decision tree ensembles on the same feature sets. The outputs of the two ensemble classifiers are finally combined with the outputs of all the other base classifiers The experimental results show the effectiveness of the above setting of multi-classifier fusion for advancing the performance in comparison with using each single (non-ensemble) learning algorithm.