Study Of Support Vector Machine Classification And Hash Retrieval Methodsbased On Structure Preserving Property

With the coming of “big data” age,the number of data categories and the size of data increase greatly,which pose enormous challenges to the tasks of data classification and information retrieval.Since the structure-preserving characteristics of data are crucial to improve the performance of classification and retrieval,this paper studies how to use the structure preserving property to improve the performance of classical methods in the field of support vector machine classification and hash retrieval.The research findings are as follows: Aiming at the problem that the inter-class and intra-class relationships in Generalized Eigenvalue Proximal Support Vector Machine algorithm(GEPSVM)are largely ignored,we propose a new method called Bounded Locality Preserving Distance based on Generalized Eigenvalue Proximal Support Vector Machine(BLPD-GEPSVM).By introducing the structure preservation matrix model,A weight matrix is defined,such that the distance between the samples within the same class of data is minimized,while the distance of different classes of data is maximized.This enables the information within the class to be kept as much as possible,and the connection between the different classes is separated in a great amount.The BLPD-GEPSVM method improves the accuracy of the GEPSVM method on cross-data and complex cross-data,as well as some benchmark UCI datasets.The experimental resuls verifies that the proposed method has better classification performance than some traditional methods.Aiming at the problem that the existing supervised discrete hash method(SDH)ignores the intraclass and interclass structure information in single-modal data,a graph-based discrete hash retrieval method(GCDH)is proposed.Based on the SDH method,the idea of graph embedding is introduced to construct the intra-class intrinsic graph,so that the distance between the same label of data is minimized.In the same time,the inter-class penalty graph is constructed,such that the distance between different labels of data has a large penalty.Further,through appropriate projection,the samples of the same class will be embedded in the subspace more compactly,and the samples from different classes are as far as possible.The experimental result demonstrate that,through the graph constraint,the discriminant information is further enhanced,thereby improving the retrieval precision.For the discrete cross-modal hash(DCH)method,the consistency in the embedded space can not be guaranteed.To solve this problem,a discrete hash retrieval method based on subspace embedding(ESCH)is proposed.This method adds an additional real embedding subspace based on the DCH model to maintain the consistency of the data pair from different modals,and then performs a second projection from the embedded space to the Hamming space.In the same time,the class label information is linked here to further enhance the discriminability.It should be noted that the proposed ESCH does not consider the data correlation within the class and between the different classes.To address this issue,we further propose an Embedding Space Consistency Structure Preserving Hashing(ESSPH)method.The DCCA model is introduced to make the same class of data have the most correlation and the correlation of different classes of data is largely prohibited.Therefore,the correlation information between different modals,as well as the intra-class and inter-class data structure in each modal,is better mined.The resulting hash codes exhibit well structure proprety and are proved to have a better performance in the cross-modal data retrieval,compared to the traditional ones.