Adversarial Deep Embedded Clustering: On a Better Trade-off Between Feature Randomness and Feature Drift

To overcome the absence of concrete supervisory signals, deep clustering models construct their own labels based on self-supervision and pseudo-supervision. However, applying these techniques can cause Feature Randomness Drift. In this paper, we formally characterize two new concepts. On one hand, takes place when a considerable portion pseudo-labels is deemed to be random. regard, trained model learn non-representative features. other Drift pseudo-supervised reconstruction losses are jointly minimized. While penalizing loss aims preserve all inherent data information, optimizing embedded-clustering objective drops latent between-cluster variances. Due compromise, clustering-friendly representations easily drifted. context, propose ADEC (Adversarial Deep Embedded Clustering) novel autoencoder-based model, which relies discriminator network reduce random features while avoiding drifting effect. Our metrics $\Delta _{FR}$ _{FD}$ allows to, respectively, assess level We empirically demonstrate suitability our handling problems using benchmark real datasets. Experimental results validate that outperforms state-of-the-art methods.