Research On Deep Clustering Based On Unbalanced Pairwise Constraints And Joint Loss

In the era of big data,clustering,as one of the most important methods in machine learning and data mining,has been widely used in many fields.Among various clustering methods,semisupervised clustering not only saves labor costs,but also obtains a relatively high-quality result by a little supervisory information.Therefore,researches on semi-supervised clustering can help many industries to analyse the large multidimensional data quickly,which has important and academic value.The supervisory information of semi-supervised clustering is mainly class labels and pairwise constraint information.In this paper,we focus on semi-supervised clustering based on pairwise constraint,and propose a Joint Loss Clustering Network(JLCN).The major contributions of this paper are as follows:1.We propose a deep clustering network with joint reconstruction loss,pairwise constraint loss and clustering loss.The deep clustering network adopts the reconstruction loss of deep autoencoder network and drives the hidden feature to learn the generality of the data.The pairwise constraint loss drives the network to learn the similarity between the data,that is,the distance of data with the same category is closer while the distance of data with different categories is further.The clustering loss can help the network to learn the similarity between the data more efficiently.Several experiments suggest that the proposed deep clustering network has a better result.2.We propose a pairwise constraint loss based on dynamic cost sensitive learning.In order to solve the problem of imbalance data in pairwise constraint information,we propose a method that enlarges the weight of the minority samples dynamically in pairwise constraint loss.The proposed method drives the network to learn the similarity between the data more efficiently and leads to a better result suggested by experiments.