Reinforced EM Algorithm for Clustering with Gaussian Mixture Models

Abstract

Methods that employ the EM algorithm for parameter estimation typically face the notorious yet unsolved problem that the initialization input significantly impacts the algorithm output. We here develop a Reinforced Expectation Maximization (REM) algorithm for cluster analysis using Gaussian mixture models. The competence of REM is achieved by introducing two innovative strategies into the EM framework: (1) a mode-finding strategy for initialization that detects non-trivial modes in the data, and (2) a mode-pruning strategy for detecting true modes/mixture components of the population. The pruning strategy is well-justified in the context of mixture modelling, and we present theoretical guarantees on the quality of the initialization. Extensive experimental studies on both synthetic and real datasets show that our approach achieves better performance compared to state-of-the-art methods