Research On Deep Multimodal Feature Clustering For Low-quality Data

In big data,there are many kinds of data with low value density and low veracity.Therefore,low quality is a significant feature of big data.Clustering is a kind of important data analysis methods.It can divide objects of data space into several classes so that data objects in the same class have similar properties while data objects in different classes have dissimilar properties.With the expansion of data scale,data dimension and data modality,traditional clustering methods have been too unsuitable for data analysis to achieve the expected performance on low-quality data.Therefore,this paper proposes solutions to the above problems,studying two deep multimodal feature clustering methods for low-quality data to effectively improve the clustering performance of clustering algorithms on low-quality,high-dimensional and multimodal data.The main contributions of this paper are as follows:(1)A multimodal fusion based Gaussian mixture model deep clustering method for high-dimensional data is proposed.First,deep neural network model was used to extract the unique features of each modality,reduce the data dimension,and concentrate the information hidden in the data.Then multimodal fusion network was used to fuse the features of each modality to obtain the fused modality features.Finally,Gaussian mixture model was used to conduct the feature clustering and obtain final clustering results iteratively.(2)A clear model supervision based multimodal deep denoising non-negative matrix factorization clustering method is proposed.The method forced deep neural networks to learn the data pattern hidden in the noisy data by adding noise to the clear training data,and the clear model was used to supervise the feature learning of each layer in the network.Finally,the non-negative matrix factorization was used to learn the clustering features with interpretability,and the low-quality data clustering task could be completed in an end-to-end manner.(3)Experiments for the two proposed methods on standard datasets were conducted.The experimental results showed that both of the two methods could capture inherent patterns in data and get the good clustering performance.Among them,the multimodal fusion based Gaussian mixture model deep clustering method for high-dimensional data showed the effective fusion ability for multimodal input data.And the clear model supervision based multimodal deep denoising non-negative matrix factorization clustering method proved the improvement of clustering effect achieved by the clear model supervision and non-negative matrix factorization.In conclusion,this paper proposes two deep multimodal feature clustering methods for low-quality data.Both of them have better clustering performance and can effectively accomplish the clustering task on low-quality multimodal data.They provide a solid foundation for further downstream data analysis tasks.