KAUST Grant NumberKUSC1-016-04
MetadataShow full item record
AbstractThis paper describes a new semi-supervised learning algorithm for intra-class clustering (ICC). ICC partitions each class into sub-classes in order to minimize overlap across clusters from different classes. This is achieved by allowing partitioning of a certain class to be assisted by data points from other classes in a context-dependent fashion. The result is that overlap across sub-classes (both within- and across class) is greatly reduced. ICC is particularly useful when combined with algorithms that assume that each class has a unimodal Gaussian distribution (e.g., Linear Discriminant Analysis (LDA), quadratic classifiers), an assumption that is not always true in many real-world situations. ICC can help partition non-Gaussian, multimodal distributions to overcome such a problem. In this sense, ICC works as a preprocessor. Experiments with our ICC algorithm on synthetic data sets and real-world data sets indicated that it can significantly improve the performance of LDA and quadratic classifiers. We expect our approach to be applicable to a broader class of pattern recognition problems where class-conditional densities are significantly non-Gaussian or multi-modal. © 2013 Elsevier Ltd. All rights reserved.
CitationYu Y, Gutierrez-Osuna R, Choe Y (2014) Context-sensitive intra-class clustering. Pattern Recognition Letters 37: 85–93. Available: http://dx.doi.org/10.1016/j.patrec.2013.04.031.
SponsorsThis publication is based in part on work supported by Award No. KUSC1-016-04, made by King Abdullah University of Science and Technology (KAUST).
JournalPattern Recognition Letters