Why does affinity propagation work so well?
Most methods that cluster by finding a set of prototypes work by keeping track of a fixed set of prototypes and iteratively refining that set. Sometimes, they make small changes to the size of the set during learning, by merging clusters or splitting clusters, but all that is kept track of is a particular configuration of the set of prototypes. In contrast, affinity propagation simultaneously considers all data points as potential prototypes and passes soft information around until a subset of data points “win” and become the exemplars. We think this gives affinity propagation an advantage over other methods.
Related Questions
- Affinity propagation seems to outperform methods that randomly resample potential centers, especially for non-trivial numbers of clusters. Why?
- To what extent can affinity propagation be carried out in real-time, i.e., when the similarities are changing in time?
- Could affinity propagation be used to obtain a hierarchical clustering of the data?
