Multi-view And Privileged Collaborative Learning Approach Study

In the era of data explosion,various technologies in the field of artificial intelligence are flourishing.Classification methods are still one of the important research fields in machine learning.The research on classification methods in machine learning has a certain practical value in specific scenarios such as image,semantic recognition and so on.With the arrival of the climax of mobile internet,data volume is increasing and data categories are enriching,and new application scenarios are constantly being proposed,which brings certain challenges to traditional machine learning methods.Classical classification methods for diverse data can not handle the internal relationship of the data well,and the generalization performance is poor.Therefore,focus on the issue of multi-view classification methods in the context of new data applications.Multi-view learning focuses on making full use of feature data obtained from different channels or levels for the same object to learn effectively.The principles of consistency and complementarity provide important guidance for multidimensional modeling.In this paper,after studying many improved multi-view learning algorithms at home and abroad,a new algorithm with strong classification performance is proposed based on privileged information paradigm and the principle of consistency between perspectives while satisfying the complementarity between perspectives,which is verified by real data sets.The main points of this paper are as follows.(1)Based on the concept of Learning Using Privileged Information(LUPI)and Random Vector Functional Link(RVFL)a classification method is proposed to monitor current perspectives by utilizing additional information from redundant perspectives on average.Fast Multi-view Privileged Random Vector Function Link Network(FMPRVFL)with two perspectives can effectively solve multi-view classification tasks.FMRVFL solves faster and provides additional generalization capabilities than classical multi-view learning methods.Our experimental results on 64 datasets show that FMRVFL is superior to the comparison method in average test accuracy and run time.(2)A Kernel Multi-view Pivileged Random Vector Function Link Network(KMPRVFL)is proposed to effectively solve multi-view classification tasks.Following the principle of FMRVFL,KMPRVFL proposes a multi-view kernelization method with two or more perspectives.The objective function of this method can find an analytic solution,we can use the pseudo-inverse method to solve it quickly,and then we carry out experiments on real multi-view datasets.Compared with common multiview classification algorithms,KMPRVFL achieves better generalization performance to prove the effectiveness of this method.For incremental application scenarios,we also propose an incremental method of KMPRVFL.Experiments show that it can effectively improve performance in incremental tasks.