Is there a probabilistic interpretation of affinity propagation?
Yes. Affinity propagation can be viewed as performing model selection and for the selected model, performing MAP estimation of the cluster centers and the assignments of data points to centers. Unlike parametric methods, the centers are forced to be on top of data points (which has advantages and disadvantages). The similarity of point x(i) to point x(k) can be thought of as -log P(data point x(i) | center x(k)) and the preference of point x(i) can be thought of as -log P(center x(i)), which corresponds to the Bayesian prior density over the center.
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?
- Is there a probabilistic interpretation of affinity propagation?
