Pseudo-Label Guided Partially View-Aligned Clustering

Many real-world data is comprised of multiple views that come from different resources, which has aroused increasing attention in recent years. Despite that existing clustering methods having achieved great performance, most of them heavily relies on the alignment of views. But in real world, it is common that only a part of the data is aligned across views due to some unexpected situations, which turning this problem into a partially view-aligned problem. To solve this challenging problem, we propose a new method which could simultaneously learn representations of data and align data from different views. The method we proposed is based on contrastive learning, while using pseudo-label to guide the aligning process. To obtain more positive pairs, we also introduce the concept of high-order neighbor, while using K-Nearest Neighbor(KNN) to gain negative pairs for contrastive learning. In brief, our method aims to find a category-level alignment relationship between different views. Since existing methods failed to make use of the clustering structure of individual view, we combine pseudo-label into this model to make better use of it. Extensive experiments on four benchmark datasets show the superiority of the proposed method over some existing state-of-the-art partially view-aligned clustering methods.