Diversity in Classifier Ensembles: Fertile Concept or Dead End?

Tree Species Classification Based on Hybrid Ensembles of a Convolutional Neural Network (CNN) and Random Forest Classifiers

In this paper, we evaluate different popular voting strategies for fusion of classifier results. A convolutional neural network (CNN) and different variants of random forest (RF) classifiers were trained to discriminate between 15 tree species based on airborne hyperspectral imaging data. The spectral data was preprocessed with a multi-class linear discriminant analysis (MCLDA) as a means to reduce dimensionality and to obtain spatial–spectral features. The best individual classifier was a CNN with a classification accuracy of 0.73 +/− 0.086. The classification performance increased to an accuracy of 0.78 +/− 0.053 by using precision weighted voting for a hybrid ensemble of the CNN and two RF classifiers. This voting strategy clearly outperformed majority voting (0.74), accuracy weighted voting (0.75), and presidential voting (0.75).

Pruning High-Similarity Clusters to Optimize Data Diversity when Building Ensemble Classifiers

Diversity is a key component for building a successful ensemble classifier. One approach to diversifying the base classifiers in an ensemble classifier is to diversify the data they are trained on. While sampling approaches such as bagging have been used for this task in the past, we argue that since they maintain the global distribution, they do not create diversity. Instead, we make a principled argument for the use of [Formula: see text]-means clustering to create diversity. Expanding on previous work, we observe that when creating multiple clusterings with multiple [Formula: see text] values, there is a risk of different clusterings discovering the same clusters, which would in turn train the same base classifiers. This would bias the ensemble voting process. We propose a new approach that uses the Jaccard Index to detect and remove similar clusters before training the base classifiers, not only saving computation time, but also reducing classification error by removing repeated votes. We empirically demonstrate the effectiveness of the proposed approach compared to the state of the art on 19 UCI benchmark datasets.