We present a novel clustering approach, that exploits boosting as the primary means of modelling clusters. Typically, boosting is applied in a supervised classification context; here, we move in the less explored unsupervised scenario. Starting from an initial partition, clusters are iteratively re-estimated using the responses of one-vs-all boosted classifiers. Within-cluster homogeneity and separation between the clusters are obtained by a combination of three mechanisms: use of regularised Adaboost to reject outliers, use of weak learners inspired to subtractive clustering and smoothing of the decision functions with a Gaussian Kernel. Experiments on public datasets validate our proposal, in some cases improving on the state of the art.
CLOOSTING: CLustering Data with bOOSTING
BICEGO, Manuele;CRISTANI, Marco;MURINO, Vittorio
2011-01-01
Abstract
We present a novel clustering approach, that exploits boosting as the primary means of modelling clusters. Typically, boosting is applied in a supervised classification context; here, we move in the less explored unsupervised scenario. Starting from an initial partition, clusters are iteratively re-estimated using the responses of one-vs-all boosted classifiers. Within-cluster homogeneity and separation between the clusters are obtained by a combination of three mechanisms: use of regularised Adaboost to reject outliers, use of weak learners inspired to subtractive clustering and smoothing of the decision functions with a Gaussian Kernel. Experiments on public datasets validate our proposal, in some cases improving on the state of the art.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
