| Multi-view learning is one of the rapid development fields of machine learning.It aims to simultaneously consider the information obtained from different angles of the same object and mine potential patterns of data according to various feature information as much as possible.Multi-view/multi-modal learning is currently being used in a wide range of AI applications,such as Autonomous Driving and Intelligent Medicine.Most of the current multi-view research methods assume that the collected data is noiseless and high-quality,on which efficient performance can be achieved.However,in practical application,data will be influenced by some cases,resulting in low-quality,such as missing data or noise pollution.The problems caused by these influences will render most current multi-view learning methods ineffective.In order to solve the problems caused by low-quality data,this paper proposes two research works to resolve data with missing instance and Gaussian noise respectively.The first model is a subspace clustering algorithm based on tensor low-rank constraint,and solve multi-view learning problems with missing views.The second model introduces gaussian process model based on neural network framework to build an end-to-end interpretable classification model.The main contributions of this paper are as follows:1.Based on tensor multi-view subspace clustering,the model in order to solve the problem of containing low-quality data with missing case,introducing hidden representation which can encode all available information,joint tensor low rank constraints simultaneously clustering and complete missing cases and to form a unified framework,thus ensuring the completion of information can be used effectively to potential model in data mining,which improve the robustness and validity of the model.2.Based on Gaussian Process multi-view classification,mainly to solve the model with modal noise on the impact of the classification task,on the basis of the neural network into the gaussian process model is used for the determination of the distribution rather than a single prediction model output value,so as to make the model can estimate the uncertainty in the result of the forecast,while containing noise uncertainty will increase rapidly,which improve the validity and interpretability of the model.3.This paper provides sufficient experimental support.In order to prove the effectiveness of the proposed algorithm model,comparative experiments of clustering/classification performance are carried out on several data sets.Detailed visualization results and proofs are provided for the convergence and complexity analysis of the model.In order to prove that the noise added by different views can be distinguished in the classification experiment with noise,the visualization result of uncertain distribution estimation is provided.The above fully proves that the proposed model has better performance and strong robustness against low quality data. |
