Semi-supervised Affinity Propagation Clustering: Algorithm and Implementation

Semi-supervised Affinity Propagation (AP) is an enhanced clustering algorithm that combines the strengths of unsupervised and supervised learning. While standard Affinity Propagation is an unsupervised method that automatically determines cluster centers and numbers through message passing between data points based on similarity measures, the semi-supervised version incorporates prior knowledge to guide the clustering process. The core innovation of semi-supervised AP lies in its two primary approaches to leveraging supervisory information: constraint integration and similarity adjustment. Constraint-based implementation allows specifying must-link constraints (requiring certain data points to belong to the same cluster) and cannot-link constraints (forcing points into different clusters). In code implementations, this typically involves modifying the responsibility and availability matrices during the message-passing phase. Similarity adjustment techniques involve preprocessing the similarity matrix using known label information to increase similarity values between same-class samples and decrease values between different-class samples, often implemented through similarity reinforcement or penalty functions. This algorithm is particularly valuable for partially labeled datasets common in applications like image segmentation and text classification. When only a small subset of samples has labels, semi-supervised AP can significantly improve clustering performance by leveraging limited supervisory information. Compared to traditional AP, the semi-supervised version typically produces more stable clustering results and more rational category divisions through its constrained optimization framework. Key implementation considerations include efficient constraint propagation algorithms and balanced similarity matrix transformations that maintain the original data structure while incorporating supervisory signals.