Fusion Methods for Unsupervised Learning Ensembles
The application of a “committee of experts” or ensemble learning to artificial neural networks that apply unsupervised learning techniques is widely considered to enhance the effectiveness of such networks greatly. This book examines the potential of the ensemble meta-algorithm by describing and testing a technique based on the combination of ensembles and statistical PCA that is able to determine the presence of outliers in high-dimensional data sets and to minimize outlier effects in the final results. Its central contribution concerns an algorithm for the ensemble fusion of topology-preserving maps, referred to as Weighted Voting Superposition (WeVoS), which has been devised to improve data exploration by 2-D visualization over multi-dimensional data sets. This generic algorithm is applied in combination with several other models taken from the family of topology preserving maps, such as the SOM, ViSOM, SIM and Max-SIM. A range of quality measures for topology preserving maps that are proposed in the literature are used to validate and compare WeVoS with other algorithms. The experimental results demonstrate that, in the majority of cases, the WeVoS algorithm outperforms earlier map-fusion methods and the simpler versions of the algorithm with which it is compared. All the algorithms are tested in different artificial data sets and in several of the most common machine-learning data sets in order to corroborate their theoretical properties. Moreover, a real-life case-study taken from the food industry demonstrates the practical benefits of their application to more complex problems.
What people are saying - Write a review
We haven't found any reviews in the usual places.
Artificial Neural Networks
Use of Ensembles for Outlier Overcoming
Ensembles of Topology Preserving Maps
A Novel Fusion Algorithm for TopologyPreserving Maps
Appendix A The Cured Ham Data Set
Appendix B Table of Experiments
Other editions - View all
AdaBoost AdaBoost.M2 Adaptation ANNs artiﬁcial average bagged ensemble Bagging algorithm Baruque calculated Cancer data set Chapter classiﬁcation accuracy classiﬁers combination computational Corchado cross-validation deﬁned diﬀerent models diﬃcult directions Distortion measure eﬀect eigenvectors ensem ensemble fusion ensemble learning ensemble meta-algorithm ensemble training Euclidean Distance experiment ﬁnal map ﬁnd ﬁrst fusion algorithms Fusion by Distance Fusion by Euclidean Fusion by Similarity Fusion by Voronoi fusion methods grid Ham data set Hebbian learning improve inﬂuence information captured Iris data set ISBN label learning algorithm Max-SIM neighbourhood neural networks neurons outliers outperform output neuron Percentage of information performed presented principal components problem quality measures obtained Quantization Error Re-Labelling Re-PCA re-sampling representation represented Self-Organizing Maps single map single model speciﬁc subsets Superposition Table tests Topographic Error topology topology-preserving maps training data training the ensemble unsupervised learning updating variance ViSOM visualization Voronoi Polygon Voronoi Polygon Similarity Weighted Voting WeVoS algorithm