Abstract:
This thesis is about a study of the behaviour of the dominant-set clustering (DS) using a measure of similarity that derives from the Euler kernel (Euler-Gauss DS), a kernel which relies on a nonlinear and robust cosine metric that is less sensitive to outliers. Moreover, in order to create a partitional clustering we use graph tranduction to propagate the membership information from the dominant sets to unlabeled data.
We perform an extensive experimental evaluation, using both synthetic and real-world datasets, in order to compare Euler-Gauss DS with the DS algorithm using the classic Gaussian kernels. Furthermore, we compare Euler-Gauss DS with other clustering algorithms, among which another method that relies on the Euler Kernel (Euler k-means).
