Semi-Supervised Affinity Propagation with Instance-Level Constraints
Inmar Givoni, Brendan Frey; JMLR W&CP 5:161-168, 2009.
Recently, affinity propagation (AP) was introduced as an unsupervised learning algorithm for exemplar based clustering. Here we extend the AP model to account for semi-supervised clustering. AP, which is formulated as inference in a factor-graph, can be naturally extended to account for ?instance-level? constraints: pairs of data points that cannot belong to the same cluster (cannot-link), or must belong to the same cluster (must-link). We present a semi-supervised AP algorithm (SSAP) that can use instance-level constraints to guide the clustering. We demonstrate the applicability of SSAP to interactive image segmentation by using SSAP to cluster superpixels while taking into account user instructions regarding which superpixels belong to the same object. We demonstrate SSAP can achieve better performance compared to other semi-supervised methods.