Flexible disentangled representation learning with soft-splitting for multi-view data

Abstract

Multi-view representation learning has gained significant attention in the machine learning and computer vision communities. However, existing approaches often fail to fully exploit the complementary part among different views during the fusion process, which may lead to representation entanglement and consequently degrade the performance for downstream tasks. To this end, we propose a novel Flexible Disentangled Representation Learning for Multi-View data in this paper. Specifically, the representation learning is performed by an adaptive soft-splitting multi-view gated fusion auto-encoder network, namely ASS-MVGFAE, which aims at separating the complementary and consistency parts in a soft way, rather than hard-splitting in the traditional methods. And then the decoupled common features are fed into a Gated Fusion Unit (GFU) to be aligned and fused, such that the shared latent representation is achieved for downstream clustering. Extensive experiments on several real-world datasets demonstrate that our method outperforms the state-of-the-art in terms of several evaluation metrics.