1: \begin{abstract}

2: In this paper we advocate the use of diffusion processes guided by

3: density to perform soft clustering tasks. Our approach interpolates

4: between classical mode seeking and spectral clustering, being

5: parametrized by a temperature parameter $\beta>0$ controlling the

6: amount of random motion added to the gradient ascent.  In practice we

7: simulate the diffusion process in the continuous domain by random

8: walks in neighborhood graphs built on the input data. We prove the

9: convergence of this scheme under mild sampling conditions, and we

10: derive guarantees for the clustering obtained in terms of the cluster

11: membership distributions.  Our theoretical results are corroborated by

12: our experiments on manufactured data and on real data.

13: \end{abstract}

14: