Research And Application On Multi-Instance Learning Using Support Vector Machine

With the continuous development of machine learning,many traditional areas use machine learning to improve research level,and the methods of machine learning application continues growing;especially the data analysis methods based on machine learning have become one of the key technologies on solving complex problems.Therefore,the role of machine learning has gradually changed.It has developed into a new stage,which a number of new methods and ideas (such as multi-instance learning,semi-supervised learning) have gradually been proposed,and the transition from theoretical analysis to practical application come true.In traditional machine learning,an object has one description,but in some practical problems,an object may have more than one description.To resolve the 1:N:1 relationship between "Object:Description:Type" is the multi-instance learning.In the multi-instance learning, training samples are bags,which are composed by a number of samples.The bag is concept marked,but the sample itself isn't concept marked.The purpose is to predict the label of a new bag.As multi-instance learning has unique character,it is considered as a new study framework apposed with supervised learning,non-supervised learning,and reinforcement learning.As multi-instance learning has unique characters and well application prospect,it has caused the extensive concern to many researchers.They have been proposed many algorithms based on multi-instance learning,and divided them into three categories.The first category specialized developed algorithm to solve multi-instance learning problems,the second category try to use multi-instance learning constraints to the objective function of single instance learning, the third category try to transform the multi-instance learning as single instance problem. Previous studies are based on the former two methods,and the research on the third method is less.Therefore,based on the third category,combining with support vector machines algorithm to improve the performance of multi-instance learning is a task worthy to research.The task of this thesis is the research of Multi-instance learning based on Supporting Vector Machine and its application.Analyze the formerly researches about Multi-instance learning methods,do some research about the existing Multi-instance learning methods based on Supporting Vector Machine,and propose three new algorithms based on the third category. Main content on the paper was summarized as follows:1.Research on multi-instance learning.Research the background of Multi-instance learning and its basic theory.Analyze the differences between multi-instance learning and traditional machine learning algorithms.Introduce the classification algorithm about multi-instance learning. At present,the multi-instance learning is widely used in pharmaceutical industry,protein recognition,image classification,content-based image retrieval,medical image-aid identification,,text classification,and stock prediction and so on.2.Research on Statistical Learning Theory and Support Vector Machine.Mainly research on the basic theory of statistical learning theory and support vector machine,and give out the basic principle of support vector machine.The improved algorithms and application of support vector machine were introduced.3.Research on Multi-Instance Learning using Support Vector Machine.Do some research about the existing Multi-instance learning methods based on Supporting Vector Machine: Multi-instance learning method based on bags(MI-SVM),Multi-instance learning method based on instance(mi-SVM),DD-SVM,MILES and so on.In this thesis,it proposed three new Multi-instance learning methods based on Supporting Vector Machine which are WEMISVM,BEMISVM and SEMISVM.It converts Multi-instance learning problems to single instance problem and achieves the outcome.The WEMISVM method used the arithmetic mean value method when predicting a new bag;the BEMISVM method used the ensemble boosting method in the integration learning;the SEMISVM method transmitted the multi-instance learning problem into single instance learning problem through calculating the arithmetic mean value to instance.4.Experiment design and results analysis.The experimental design is done on the weka platform,using 10-fold cross-validation methods,and uses the evaluation standards consist of accuracy of classification,root mean square error and Kappa statistical value.Selected 14 groups experimental data sets which are provided by Professor Eibe Frank.Compared to the standard Multi-instance learning algorithms(CitationKNN,MDD,MIDD,MIEMDD,MINND,MILR,MISMO,MISVM) and traditional machine learning algorithms under three learning mechanisms, we can verify the performance of the new algorithms,and make a comparison and analysis.Experiment results show that the three new Multi-instance learning methods based on Supporting Vector Machine perform well;they have good application prospects.