Research On Key Technologies Of Clustering Based On Deep Learning

Nowadays, deep learning is one popular topic in artificial intelligence and machine learning. And deep learning shows great capability at unsupervised feature learning field. Deep Learning is a general term of a set of algorithms which learn the representative features about the real word data sets with deep architectures.Clustering method is the hot topic of machine learning and data mining, and it is also one unsupervised learning method. Clustering is learning the hidden cluster structure in the dataset with no prior knowledge.In this thesis, we propose a novel clustering model combined deep architecture and clustering method. And we introduce a novel clustering algorithm deep belief networks oriented clustering (DBNOC) combining deep belief networks and fuzzy c-means(FCM). The proposed model divides into two parts:pre-training part and further tune part. The pre-training part contains multi RBM and one pre-clustering layer. There are two sections in the pre-training:one is that we use deep architecture to learn the high level representation. The other is that we using one kind of simple clustering algorithm clustering in the high level representation space to produce the initial cluster center vector. Afterwards we using the tune network and center holding clustering method to optimize the loss function in a cross iteration way. The new model takes advantage of both deep learning and clustering to make clustering more simple and more convenient. The clustering model based on deep learning aims at learning deep feature representation when clustering and finding the hidden cluster structure from the deep feature representation. The newly proposed algorithm fulfills the model with deep belief networks and FCM algorithm. The DBNOC algorithm not only takes the advantage of deep belief networks, also is a fuzzy clustering method.Finally, we do experiments on both high dimension datasets and low dimension datasets. The experiment results suggest that the clustering based on deep learning model and DBNOC algorithm is better than the standard k-means and FCM algorithm on the test datasets. And it shows more generalization. We discuss the parameter in this algorithm. We also discuss the algorithm performance under different value of learning rate, regulation factor. What is more, the proposed model and algorithm is suitable in theory and practice.