Research About Distance Metric Based On Multi-view Subspace Learning

Nowadays,with the improvement of the technology of data collection equipment,there are more and more kinds of data in different fields.This results in some algorithms that rely on a given metric being inapplicable to data in specific contexts.For example,KNN,Kmeans,etc.,use the given Euclidean distance to calculate the distance from the sample point to the center,which often ignores the diverse nature of the data.The proposal of Distance Metric Learning(DML)occupies an important position in the field of pattern recognition,especially in image recognition,with breakthrough results.As a classic distance metric learning method,Distance Metric Learning for Large Margin Nearest Neighbor(LMNN)algorithm has also attracted widespread attention from researchers.With the advent of the era of big data,data has gradually changed from single lightweight data to heavy-weight multi-view data.Multi-view data for different descriptions of the same object has spread in various fields.For example,they use a drone to capture the same data set from multiple angles.However,most existing DML methods cannot be applied to multi-view data.They cannot guarantee independent and shared feature subspaces from multiple sources or different feature subsets,thus ignoring statistical feature information in model learning.Researchers target Multi-view Distance Metric Learning(MVDML).This paper proposes a series of novel multi-view metric learning methods based on the idea of a large margin.It simultaneously validates the effectiveness of the forms on remote sensing,face,forest fire,person,and UCI datasets.The innovative achievements of this paper mainly include:1)This thesis proposes a new DML model called Multi-view Distance Metric Learning Via Independent and Shared Feature Subspace.It learns multiple distance metrics to unify information from various views.The method finds a distance metric for each view in independent feature space to keep its specific properties.It also finds a distance metric-dependent sparse representation for different views in shared feature space to keep their common Attributes.MVML-ISFS is based on LMNN to formulate a multi-view large margin loss function,which encourages each view to have a large separation so that the distance between each sample of the same class pair is smaller than the distance between samples of each class pair.The model involves multiple variables,and the gradient descent strategy is used to optimize the multiple variables.2)In order to further explore the potential correlation between views,this thesis proposes a new DML model based on MVML-ISFS,called Multi-view Metric Learning for View Cross Discrimination.The model introduces a regular term to MVML-ISFS,which maximizes the difference between different types of samples between different views while ensuring the similarity of the same samples between different views,aiming to improve the classification performance further.The algorithm involves multiple variable parameters,so the algorithm is solved by using the gradient descent algorithm to optimize the multiple variables.3)This thesis conducts extensive and extensive experiments on UCI and several application problems with KNN as the evaluation standard,such as face recognition,forest fire recognition,person recognition,and remote sensing classification.Experimental results show that our proposed MVDML method outperforms most state-of-the-art DML methods,implying that these algorithms can serve as an effective classification tool.