Research On Improved Algorithm For Multi-instance Learning MILES

Multiple-Instance Learning(MIL)is a variant of a traditional learning setting,MIL has been widely studied as it applies to text categorization;image processing(natural scene classification,content-based image retrieval,image categorization),etc.In order to solve many kinds of MIL problems in recent years,researchers have proposed a variety of algorithms.Where the classical embedded space algorithm MILES use of package level mapping to construct the packet level features,then use SVM to train a standard supervised learning classifier.MILES mainly divides into three parts: feature selection and feature mapping,as well as standard classification structure.A lot of experimental data sets show that the algorithm classification accuracy is high.But because of a vast quantity of data processed,causing MILES of high computational complexity,the algorithm's running time is long,the anti-noise performance of MILES also needs to be improved.Therefore,using the ideas of MILES to make new and improved others,the main contents are as follows:(1)MILES and other typical algorithms are analyzed.Especially analyzed several typical MIL algorithms,such as DD,DD-SVM,MILES and so on.MILES algorithm(Multiple Instance Learning via Embedded Instance Selection)is analysised in detail.The general idea of this algorithm is to transfer MIL problem into an existing supervised learning problems.Specifically,using all single example of the training set to construct a feature space.Then do feature selection.Finally,use the selected eigenvectors to construct a standard training classifiers.This article will analyze the whole idea of MILES and the two key issues:feature selection and feature mapping,as well as standard classifier construction.Then combining with the experiment result,analyse the shortcomings of MILES algorithm,make bedding for proposing improved algorithms.(2)Due to MILES's high algorithm complexity,and its anti-noise performance is also not good enough,a new improved algorithm MIL-CPNN(Multiple Instance Learning via Probabilistic Neural Network based on Clustering)is proposed.In this paper,make use of the general idea of MILES,we first get a package level characteristics,then construct a classifier.In order to avoid a lot of computation,first obtain packag level features by clustering,Then make use of probabilistic neuralnetwork's efficiency,send the features into the trained PNN to construct a classifier.The experimental results show that this algorithm has a good anti-noise performance,can effectively reduce the running time,also has a good classification accuracy.(3)In order to further improve the classification accuracy of algorithm for MILES,decrease the amount of MILES in the calculation of the embedding process at the same time,then another form of MIL algorithm which based on Embedding and Neural Network is proposed.Because it still contains embedding,while combining the neural network,so we name it MIL-ENN(Multiple Instance Learning via Embedding and Neural Network).The experimental results show that this algorithm can improve the classification accuracy to some extent,besides,the embedded space's computation complexity is reduced at the same time,hence the computational complexity of this algorithm also be effectively reduced by comparison.