Multi-task self-supervised learning based fusion representation for Multi-view clustering

Multi-view data clustering aims to utilize the complementary information from different views and is widely studied in unsupervised learning. However, it is a significant challenge to extract the common and view-specific representations within a single framework. Concurrently, Multi-task Learning (MTL) excels at learning common and specific representations from multiple inputs, but it is primarily designed for supervised learning. Direct applying of existing MTL training methods to unsupervised scenarios is not feasible due to the lack of labels. To address these challenges, we propose a Multi-task Self-supervised Learning based Multi-view Fusion Representation (MSLMFR) model, which adapts the MTL frameworks to unsupervised scenarios, with three types of tasks specifically designed. Distinct from conventional MTL frameworks, our main tasks focus on the fusion representation and reconstruction of different views. Additionally, an auxiliary anchor task is introduced to further enhance performance. Subsequently, we propose a multi-task gradient optimization method. This method updates the gradient through matrix alignment in both parameter and task spaces, and takes a norm for each task. We conduct experiments on six real- world datasets, including both images and texts. The results demonstrate the effectiveness and efficiency of MSLMFR in learning high-quality fusion representations.