Wasserstein Adversarial Mixture Clustering

Warith Harchaoui (1) , Andrés Almansa (1) , Pierre-Alexandre Mattei (1) , Charles Bouveyron (1)

1 MAP5 - UMR 8145 - Mathématiques Appliquées Paris 5

Clustering complex data is a key element of unsupervised learning which is still a challenging problem. In this work, we introduce a deep approach for unsupervised clustering based on a latent mixture living in a low-dimensional space. We achieve this clustering task through adversarial optimization of the Wasserstein distance between the real and generated data distributions. The proposed approach also allows both dimensionality reduction and model selection. We achieve competitive results on difficult datasets made of images, sparse and dense data.

Mots clés: Generative Adversarial Networks; Clustering; Deep Generative Models; Wasserstein Distance

Domaines: Machine Learning [stat.ML]

Warith Harchaoui, Andrés Almansa, Pierre-Alexandre Mattei, Charles Bouveyron. Wasserstein Adversarial Mixture Clustering. 2018. ⟨hal-01827775v2⟩