Adaptive View-Aligned and Feature Augmentation Network for Partially View-Aligned Clustering

Abstract

As an important task in multi-view clustering, partially view-aligned clustering has attracted increasing attention in recent years. However, previous algorithms have two limitations: (1) they manually calculate the fixed alignment matrix based on Euclidean distance and use the fixed matrix for common feature expression. The manual fixed alignment matrix fails to adequately reflect the similarity of the training data; (2) the process of learning features is isolated from the downstream clustering task, thus learned features are unsuitable for the clustering scenario. In this paper, we propose an adaptive view-aligned and feature augmentation network (AFAN) to tackle these two issues. First, we propose an adaptive view-aligned module to calculate the alignment matrix with the self-attention mechanism. The calculated alignment matrix can capture data similarity by jointly learning data features and view alignment. Second, we introduce a self-augmentation strategy to encourage the learned features of the same cluster to be crowded together. Extensive experimental results show that AFAN outperforms state-of-the-art approaches on four benchmark datasets.

This work was supported in part by the National Natural Science Foundation of China under Grant 61972065, Grant 62006034; in part by the Natural Science Foundation of Liaoning Province under Grant 2021-BS-067; in part by the Social Science Planning Foundation of Liaoning Province under Grant L21CXW003; in part by the State Key Laboratory of Novel Software Technology, Nanjing University under Grant KFKT2022B41; and in part by the Dalian High-level Talent Innovation Support Plan under Grant 2021RQ056.