تركيب يادگيري متري بدون ناظر و خوشه بندي: يادگيري هاي فازي و غيرخطي

چكيده انگليسي :

AbstractThe learning process in lots of machine learning algorithms is based on the similarity of the training data Somekinds of specific similarity measures for two different categories of structural and vector data exist In manycases Euclidean distance which is applied for the vector data is not a suitable criterion for determining thesimilarity among input data In distance metric learning the new projected space is extracted in such a waythat the new Euclidean distance well illustrates the similarity of the data A considerable number ofunsupervised and supervised metric learning algorithms exist In unsupervised metric learning algorithms thelinear or nonlinear subspace of the data is learned regardless of the data class labels Supervised algorithmsrequire a collection of labeled training data In this category the objective is to find the best representation ofdata through which the class separation is maximized In this dissertation we focus on the unsupervised metric learning with real world application in clustering Inunsupervised clustering the objective is to find a set of clusters where each cluster only contains the sameclass data while compactness and separability on all of them are well defined Without the label information finding such well defined clusters is impossible However one way to achieve better results is to first apply adimensionality reduction algorithm and then cluster data in the low dimensional space Another way is to firstperform a clustering method in the input space to provide an estimation of the class labels They are then fedinto a supervised metric learning algorithm to learn the better metric through which the estimated classseparation is maximized In this approach the first clustering does not use the benefit of data s underlyinglower dimensional manifold and can mislead the other steps A better way is to perform both clustering andmetric learning together which is the focus of this dissertation In this dissertation two different linear and nonlinear algorithms are introduced in this area Dis FCM is alinear method in which unlike most available methods that apply k means clustering a new formulation ispresented in which the benefits of FCM clustering are applied for obtaining the estimated class labels In thisalgorithm two closed form formula are presented for updating the clustering parameters and a SDPoptimization problem is presented to calculate metric learning parameters Thus the implementation of thisalgorithm is very simple and it will be possible to calculate the time complexity of this method MGP LVMiand its fast version are the probabilistic and non linear algorithms that in contrast to the Dis FCM apply thecrisp clustering The probabilistic model presented in the MGP LVMi provides the dimensionality reductionwhich is able to capture the latent space in appropriate manner The time complexity of the MGP LVMi issignificantly reduced compared to the Dis FCM algorithm However the improvement of its time complexityis one of the focus of this research The fast version of MGP LVMi is named as the MFGP LVM algorithm inthis dissertation The effectiveness of these methods is assessed empirically through different experiments The clustering accuracy and Normalized Mutual Information NMI are applied as the two criterions forcomparing different methods Both clustering accuracy and NMI are extrinsic measures and assume theknowledge of the ground truth Without exploiting this knowledge to define the compactness and separationof the learned clusters different intrinsic measures are applied We use different demonstrative examples toshow the effectiveness of our proposed methods in combining of metric learning and clustering Theexperimental results are also implemented on different datasets of UCI repository On these datasets the Dis FCM can improve the accuracy criterion by 5 67 the MGP LVMi by 2 89 and MFGP LVM by 2 09 over other state of the art algorithms Keywords Machine Learning Pattern Recogni

مصلحي، زهرا

تركيب يادگيري متري بدون ناظر و خوشه بندي: يادگيري هاي فازي و غيرخطي

يادگيري ماشين , شناسايي الگو , يادگيري متري , خوشهبندي , مدلهاي گرافيكي احتمالي , فرآيندهاي گوسي