Deep Clustering Based on Contractive Autoencoder and Self-paced Learning
HaoRan Bu · 2023
Cluster analysis is a common analysis method in the field of data mining, which aims to partition unlabeled input data into different clusters through a predefined similarity measure. With the advent of the era of big data, people often face the problems of large scale and high dimensional sample data. At this time, using the traditional clustering algorithm may lead to problems such as long time consumption and poor clustering performance. Therefore, this paper proposes a deep clustering algorithm based on contractive autoencoder and self-paced learning. In the pre-training stage, robust features are learned by using a trained contractive autoencoder, and the Frobenius norm is added as a penalty term in the conventional autoencoder framework to avoid learning meaningless features. In the tuning phase, the clustering loss is directly applied to the learned features, which are jointly refined and assigned to clusters. At the same time, a self-paced learning mechanism is introduced to stabilize the training process, and the most confident samples are selected in each iteration. Finally, the effectiveness of the proposed method is verified by comparing with domestic and foreign deep clustering algorithms on four commonly used data sets.