Ensemble Clustering via Learning Representations from Auto-encoder

Representation learning is central to clustering analysis and visualization. However, linear dimension reduction methods cannot model high dimensional and sparse feature spaces, thus may neglect some important information during feature representation. In this paper, we focus on developing effective models to capture non-linear and most salient features for clustering analysis, and propose an ensemble framework ECLRA to fuse different layer representations from auto-encoders. To further improve the clustering performance and fully realize the potential of ensemble learning, we also design a strategy to generate enough basic partitions, which turns out to be very effective. Experimental results demonstrate that ECLRA can significantly improve clustering performance compared with baseline methods, and are fairly robust on imbalanced datasets.