Is damping needed? How does it affect the final result?
Usually, yes. Loopy belief propagation (on which affinity propagation is based) can be viewed as a particular kind of over-relaxation. Damping is commonly needed in over-relaxation methods and here it prevents the availability and responsibility updates from overshooting the solution and leading to oscillations. As long as affinity propagation converges, the exact damping level should not have a significant affect on the resulting net similarity. Higher damping factors will lead to slower convergence.
