Research On Deep Clustering With Pairwise Constraints

Cluster analysis is an unsupervised learning method,which makes objects have high intra-cluster similarity and low inter-cluster similarity.Cluster analysis is much more difficult than supervised learning(classification).In recent years,deep neural networks(DNNs)have achieved great success in the fields of image classification,natural language processing and speech recognition.One of the reasons why DNNs have achieved great success is that they can automatically extract various hierarchical features data.In this paper,we use deep neural networks and weak-supervised learning – pairwise constraints – to cluster.The pairwise constraints are binary relationships between objects,that is,two objects must belong to the same cluster(must-link),or two objects can not belong to the same cluster(cannotlink).Compared with the class tag information,these pairs of constraints are easier to obtain.Aiming at the possible shortcomings of the existing constrained clustering models and deep clustering models,this paper proposes a deep clustering model that integrates pairwise constraints,which includes a fully-connected neural network embedded in the clustering algorithm.Then,for given pairwise constraints,we use them to train the fully-connected neural network.Finally,the neural network is used to predict the twoelement relationship of the sample pair,and then the clustering results are obtained by combining with the K-means algorithm.Although we used k-means and a fullyconnected neural network in this paper,the existing distance-based clustering algorithm and other neural networks can also be applied to the proposed clustering framework.Experimental results show that the model has good clustering effect on multiple datasets,which verifies the validity of the model.The main contributions of this dissertation are as follows:1)Proposes a deep binary classifier that learns the relationships of pairwise constraints.Trains the binary classifier with the pairwise constraints of the input.At the same time,design experiments to verify that the predicted pairwise constraints by the deep binary classifier are reasonable,that is,most of the prediction pairwise constraints are consistent with the actual pairwise constraints.2)Proposes a deep clustering model that integrates pairwise constraints.The model uses the K-means algorithm combining with the deep binary classifier described above to cluster.At the same time,the neural network part of the model can be used to predict the relationships between samples and cluster centers,and samples will be assigned to the reasonable cluster.3)Systematically compare the proposed deep clustering model with the existing clustering algorithms to verify the effectiveness of the proposed model.