Application of Machine Learning

The Goal Of This Book Is To Present The Latest Applications Of Machine Learning, Which Mainly Include: Speech Recognition, Traffic And Fault Classification, Surface Quality Prediction In Laser Machining, Network Security And Bioinformatics, Enterprise Credit Risk Evaluation, And So On. This Book Will Be Of Interest To Industrial Engineers And Scientists As Well As Academics Who Wish To Pursue Machine Learning. The Book Is Intended For Both Graduate And Postgraduate Students In Fields Such As Computer Science, Cybernetics, System Sciences, Engineering, Statistics, And Social Sciences, And As A Reference For Software Professionals And Practitioners. The Wide Scope Of The Book Provides Them With A Good Introduction To Many Application Researches Of Machine Learning, And It Is Also The Source Of Useful Bibliographical Information. This chapter discusses the issues faced when trying to train SVM on imbalanced datasets. The main reason why SVM performs poorly for such datasets is because of the weakness of soft margins. Soft margins were introduced in order to make SVM resilient against nonseparable datasets. The idea was to tolerate some classification error as a trade off for maximizing the margin between the support vectors. But this has an adverse effect when it comes to imbalanced datasets. SVM ends up with a hyper plane that is far from the entire cluster of instances. Any instance that is on the same side of the plane as the cluster is classified as the majority class. Having the hyper plane further from the instances maximizes the margin, at the cost of misclassifying the minority class. But since there are only a few instances of the minority class, the error is rather small and the benefit of larger margins overcomes this. Therefore, everything is classified as the majority class. Some solutions to this problem are presented in the chapter and evaluated against human genome and network intrusion datasets. The imbalance ratio in the genome dataset can be as high as 1:4500. This is well beyond the capability of traditional ML algorithms. However, we can use under sampling of the majority class and heuristics to reduce the imbalance ratio. Then we can use the techniques presented in this chapter to improve the performance of SVM on this dataset. These techniques include generating and selecting good features, using different error costs for majority and minority instances, and generating synthetic minority instances to even out the imbalance. Then we discussed datasets where the imbalance ratio is not known at the time of training. Network intrusion is one such application domain where the number of malicious nodes in the network is unknown at the time of training. We introduced dynamic thresholds to try to estimate this proportion and then adjust the SVM model's parameters to significantly improve its performance. We also showed that building Reputation Systems and automatically determining their rule sets using Machine Learning is not only feasible, but yields better results than some of the manually generated rule sets found in the literature. Although the techniques presented in this chapter have been shown to significantly improve SVM's performance on imbalanced datasets, there are still limitations on what degrees of imbalance SVM can handle. We have tested SVM on imbalance ratios of 1:100, however, bioinformatics datasets have imbalances of 1 to several thousands. In future, researchers need to invent better algorithms that are capable of handling such huge imbalances In this work, we report experimental results about the classification of Internet traffic by examining the packet flow in the client-server direction. We focus on the problem of early application identification, which requires to find a balance between the classification accuracy and the number of packets required by the classifier. The contribution of this work is twofold. First, we compare the performance of some wellknown supervised and unsupervised classification techniques, namely the C4.5 decision tree, the Support Vector Machines, and the Simple K-Means. We performed validation tests on three traffic traces containing a mix of traffic from different applications and concluded that the C4.5 decision tree algorithm has the best performance overall, even if the SVMs follow closely. The unsupervised technique always yields the worst performance. Second, we introduce a new classification scheme based on the observation that the connection generation process from a given traffic source is influenced by the application generating the requests. In particular, we show that, in experimental data, the Power Spectral Density of such processes often shows a power-law behavior. Therefore, we propose to use the measured power-law exponent of the traffic source as an additional feature in the classification of a traffic flow. This new feature comes at no additional delay, because its computation is based on the timestamps of the initial packets of past flows. By using this feature we were able to significantly reduce the classification error rate in all the considered scenarios. Further, we also propose an enhanced scheme in which we perform classification using the first 5 packets in a flow for low-activity sources and the first 3 packets in flow for high-activity sources. By using this scheme, we obtain a low error rate and, at the same time, we have low average classification delay. There are some possible future directions for this research. In this work, we did not consider the problem of training a classifier on a data set collected on a given link and used on a different link. We expect that the classification error rate increases, but that the per-flow features still yield an increased accuracy, because the connection request process mainly depends on the application and is weakly dependent on the specific network context. In order to study the portability of a classifier it is necessary to use traces captured at different sites but in the same day and at the same hour. This is because traffic patterns evolve over time. Another possible future work is the study of the temporal evolution of the connection generation process Multi-Scale Modeling and Analysis of Left Ventricular Remodeling Post Myocardial Infarction: Integration of Experimental and Computational Approaches