A0980
Title: Convex non-convex clustering with generalized Moreau envelope
Authors: Zheming Gao - Rice University (United States) [presenting]
Eric Chi - Rice University (United States)
Abstract: Convex clustering has attracted significant attention in unsupervised learning. A novel convex nonconvex clustering model is proposed incorporating a generalized Moreau envelope (GME) penalty. Despite the nonconvexity introduced by the GME penalty, an efficient parameter selection condition is established that ensures the overall convexity of the model. To solve the proposed formulation, a fast forward-backward splitting algorithm is developed. Numerical experiments on synthetic and benchmark datasets demonstrate the model's effectiveness, showing advantages in clustering paths and key clustering metrics, including adjusted rand index (ARI) and variation of information (VI). The proposed model provides reasonable and dominant clustering results in breast cancer subtype detection with TCGA data.
