A self-adjusting spectral clustering algorithm

Abstract

We introduce a new similarity measure for spectral clustering which produces an algorithm with input given by a data set and by the desired number of clusters. This is different from traditional similarity measures for spectral clustering in the literature, which rely on additional scaling parameters that have to be adjusted according to the dataset. Consequently, these other algorithms require considerable user expertise regarding both the chosen method and the data being analyzed. In contrast, our method does not require adjusting any scaling parameters, making our algorithm more accessible to users from various fields. Our experiments showed that this algorithm performs very well on synthetic data sets having complex shapes and multiple scales. For real data sets, the results were very competitive in comparison with traditional spectral algorithms.