(A) comparison study on classification performance using support vector machine

김은주 Graduate School, Yonsei University 2002 국내석사

전통적인 판별·분류 (discriminant and classification) 문제는 Fisher의 판별분석이나 로지스틱 회귀분석과 같은 기법들이 이용되어 왔다. 이 논문에서는 최근 Vapnik이 고안해낸 support vector machine (SVM) 방법이 분류와 회귀분석 문제에 있어서 뛰어난 성능을 보이고 있다고 알려지고 패턴인식 (pattern recognition) 방법, 특히 face recognition 문제에 널리 쓰이게 됨에 따라서 이 방법을 통계학의 판별·분류분석에 적용하여 보고자 하였다. 이 논문은 기존에 쓰이던 Fisher의 판별분석이나 로지스틱 회귀분석의 분류성능과 support vector machine의 분류성능을 오분류율로 비교하고, 궁극적으로는 support vector machine의 우수성을 확인하는 데에 그 목적이 있다. 이를 위해 여기에서 사용된 자료들은 그 크기가 비교적 작기 때문에 오분류율을 추정하기 위한 방법으로 train-and-test나 crossvalidation을 이용하는 것은 적당하지 않았다고 보아 bootstrap 방법을 이용하였고, 이 때 bootstrap 반복의 횟수는 200회로 하였다. 이와 같이 Fisher의 판별분석, 로지스틱 회귀분석, support vector machine 방법을 간단한 몇 가지 자료들에 적용하여 본 결과, support vector machine 방법의 분류성능이 다른 방법들에 비해 우수함을 나타내었다. As classical methods of discriminant and classification analysis, Fisher's discriminant analysis and logistic regression have been widely used. In this thesis, support vector machine (SVM) which was proposed by Vapnik in 1995 is introduced and applied to classification problem. This paper aims at comparing classification performances of three mentioned methods and confirming the better classification performance of SVM. Several data sets used here were so small and simple that it was not sufficient to compare only their apparent error rates or to use train-and-test or crossvalidation. Therefore, the bootstrap method was used for estimating error rates and the number B of bootstrap replicate samples was 200. Discriminant analysis, logistic regression, and SVM for classification problem were applied to the several simple data sets and the error rates of SVM were smaller than others, that is, SVM performed better than others methods.

Data Classification with Support Vector Machine

김재홍 부산대학교 대학원 2018 국내석사

In modern society, there are many unlimited amounts of data. For example, if a human being is able to perceive and observe from a minor phenomenon to a large phenomenon, this becomes data. Analyzing this data is very difficult, but very important. Classification during data analysis is a very basic method, but its effect is certain. If we have the criteria for classification, we will be able to know the meaning of the data in a complex data pattern just by satisfying or not satisfying the criteria when analyzing such data. If the criterion of the classification is set mathematically, the criterion can be used as a method in data analysis. The most popular method in such a mathematical classification method is a support vector machine. In this paper, we aim to study various kinds of support vector machines. In this paper, we will analyze the difference between the various methods of the support vector machine, and we consider that what kind a reason make us to use the support vector machine when we treat data. Finally we consider how to use support vector machine for more efficient results. iv 현대 사회에서는 데이터가 무궁무진하게 많이 생성된다. 예를 들어 사소한 현상에서부터 큰 현상까지 인간이 인지하고 관측가능하다면 이것은 데이터가 된다. 이러한 데이터를 분석하는 것은 매우 어렵지만, 매우 중요한 부분이다. 데이터 분석중에 분류는 아주 기초적인 방법이지만, 그 효과는 확실하다. 분류의 기준을 우리가 가지고 있다면, 이후에 그러한 데이터를 분석할 때 그 기준을 만족하느냐 만족하지 못하느냐만으로도 복잡한 데이터의 양상에서 데이터의 의미를 어느정도는 알 수 있게 된다. 그 분류의 기준을 수학적으로 타당하게 설정한다면 그 기준이 데이터 분석에서 활용가능한 방법이 되게 할 수 있다. 그러한 수학적인 분류 방법에서도 가장 많이 활용되는 방법이 서포트 벡터 머신이다. 본 논문에서는 다양한 서포트 벡터 머신의 종류에 대해서 연구하는 것을 목적으로 한다. 서포트벡터머신의 여러가지 방법중에 각각의 차이점을 분석하고, 어떠한 이유에서 데이터를 분석할 때에 서포트 벡터 머신을 사용하는지 알아볼 것이다. 그리고 더 효율적으로 서포트 벡터 머신을 사용할 수 있는 방법에 대해서 알아본다.

Instance-based Entropy Classifier for Imbalanced Classification Problem

조풍진 서울대학교 대학원 2019 국내박사

클래스 불균형 데이터를 바탕으로 한 지도학습은 많은 분야에서 중요한 문제로 여겨져 왔다. 소수 데이터의 무시로 인해 일반적인 분류 알고리즘과는 다른 방법이 필요하기 때문이다. 이러한 맥락에서, 퍼지 서포트 벡터 머신(Fuzzy Support Vector Machine, FSVM)은 클래스 불균형 데이터를 처리하기 위해 각 데이터 포인트의 가중치를 다르게 할당할 수 있으며, 가중치를 결정하는 연구들이 활발하게 수행되었다. 그러한 방법들 중에서 정보 이론의 엔트로피는 데이터의 설명력을 가지고 있기 때문에 퍼지 서포트 벡터 머신에 적용 할 수 있다. 또한, 클래스 불균형 분류에 대한 정보의 확실성을 정량화하기 위해 최근점 이웃의 클래스에 기반한 최근접 이웃 엔트로피 개념이 제안되었다. 그러나 기존의 엔트로피 퍼지 서포트 벡터 머신(Entropy Fuzzy Support Vector Machine, EFSVM)은 모델을 학습할 때 통일된 이웃 크기를 사용하여 오분류를 유발한다. 그래서 이 논문은 이웃의 클래스를 보다 잘 반영하는 새로운 사례 기반 분류기를 개발하는 것을 목표로 한다. 먼저, 제안된 사례 기반 엔트로피 퍼지 서포트 벡터 머신(Instance-based Entropy Fuzzy Support Vector Machine, IEFSVM)은 최근접 이웃 엔트로피의 그래프 패턴을 기반으로 개발되었다. 고정된 데이터 포인트에 대해 엔트로피 값이 이웃 크기에 따라 달라질 수 있다는 것을 참고한다면, 여러 이웃 크기에 따른 엔트로피 조합을 고려할 수 있다. 그리고 그 엔트로피 조합의 그래픽 패턴을 사용하여 합리적인 추론을 통해 가중치를 할당한다. 두 번째로, 공공 데이터와 실제 데이터를 사용하여 여러 벤치마크를 통해 IEFSVM의 성능을 입증한다. IEFSVM의 기본 분류기는 서포트 벡터 머신(Support Vector Machine, SVM)이기 때문에, 벤치마크를 구성할 때 SVM을 기본 분류기로 사용하는 알고리즘과 그렇지 않은 알고리즘 두 가지를 사용한다. 특히, 제안된 IEFSVM은 EFSVM을 포함한 다른 벤치마크들보다 높은 수신자 조작 특성 곡선의 밑 면적(Area Under the receiver operating characteristic Curve, AUC)값을 가지며 통계적으로 개선된 예측 성능을 보여준다. 마지막으로 Peer-to-peer(P2P) 대출 시장에 IEFSVM 모델을 적용하여 투자 의사 결정 모델을 개발한다. P2P 대출 시장에서 대출 상태는 불균형한 데이터이기 때문에 IEFSVM을 적용하면 완납된 대출을 예측할수 있다. 또한, 수익성을 높이기 위해 다중 회귀 분석 모델을 사용하여 높은 투자 수익을 가지고 파산하지 않을 대출을 찾는다. 흥미롭게도 IEFSVM은 분류 성능 측면에서도 기존의 클래스 불균형 분류기를 개선하고, 수익성 성과와 관련하여서도 투자 의사 결정 모델을 개선하는 데에 성공한다. 결론적으로, 이 논문의 기여도는 새로운 비용 민감 분류기의 개발과 수익성 있는 투자 결정을 위한 분류기의 응용을 포함한다. Imbalanced classification, a supervised machine learning with class imbalance datasets, has been a significant problem in many areas. Due to the ignorance of minority data, a method different from the standard classification algorithm is needed. In this context, fuzzy support vector machine (FSVM) can assign the weight of each data point differently to handle the imbalanced datasets, and the studies in determining the weight have been actively conducted. In information theory, entropy possesses a descriptive power of data, and it can be employed to FSVM. To quantify the certainty of information for imbalanced classification, nearest neighbors entropy, an entropy value based on the neighbors' class, is proposed. However, the existing entropy fuzzy support vector machine (EFSVM) employs a unified neighborhood size when learning the model, which causes misclassification. That's why this dissertation aims to develop the new instance-based classifier which better reflects neighbors' class. At first, the model of proposed instance-based entropy fuzzy support vector machine (IEFSVM) is developed based on the characteristics of nearest neighbors entropy. Given that the entropy of a fixed data point can vary according to neighborhood size, the entropy combination with several neighborhood sizes can be considered. Then, the graphical pattern of entropy combination is employed for assigning the weight with rational reasoning. Secondly, the model of IEFSVM is validated using public and real-world datasets with several benchmarks. Since the base classifier of IEFSVM is support vector machine (SVM), the benchmarks for comparison are twofold: algorithms using SVM as the base classifier and those not. Specifically, the proposed IEFSVM exhibits the statistically improved prediction performance with higher area under the receiver operating characteristic curve (AUC) than other benchmarks including EFSVM. Lastly, the model of IEFSVM is applied into Peer-to-peer (P2P) lending market to develop an investment decision model. Since the loan status of borrowers in P2P lending market is an imbalanced data, applying IEFSVM can predict fully paid loans. To enhance the profitability, a multiple regression model is also generated to detect non-default loans with high investment return. Interestingly, IEFSVM succeeds to improve the existing imbalanced classifier in terms of classification performance and even to develop an investment decision model with respect to profitability performance. In conclusion, the contribution of this dissertation involves the development of a novel cost-sensitive classifier and the application of classifier to profitable investment decision.

Ranking Support Vector Machine Visualization Using Nomogram

Nguyen Thi Thanh Thuy 경희대학교 2009 국내석사

In this thesis, we propose a visualization model for a trained ranking support vector machine using nomogram. In addition, we introduce a feature selection method for the ranking support vector machine, and show visually each feature's e??ects on the log odds ratio on the nomogram. Nomogram is a well-known visualization model that graphically describes the complete model on a single graph. The complexity of the visualization does not depend on the number of the features, but on the properties of the kernel. The experiments will show the e??ectiveness of our proposal which helps the analysts study the effects of predictive features. This method also displays its robustness in eliminating irrelevant and redundant features, then improve the overall accuracy. To represent the e??ect of each feature on the log odds ratio on the nomograms, we propose a probabilistic ranking support vector machine function. Recently, Support Vector Machines (SVMs) have been applied very e??ectively in learning ranking functions (or preference functions). They intend to learn ranking functions with the principles of the large margin and the kernel trick. However, the output of a ranking function is a score function which is not a posterior probability which can e??ectively visualize ranking support vector machine. One approach to deal with this problem is to apply a generalized linear model with a link function and solve it by calculating the maximum likelihood estimate. But, if the link function is nonlinear, maximizing the likelihood will face with difficulties. Instead, in this thesis we propose a new approach which trains an SVM for a ranking function, then maps the SVM outputs to a probabilistic sigmoid function whose parameters are trained by using cross-validation. Proposed function will be evaluated the accuracy on two data mining datasets (synthetic and OHSUMED datasets) and compared to the results obtained by standard ranking SVMs that is the most favorite ranking method now.

Discovering communities in social network service using the machine learning mixture

민무홍 성균관대학교 일반대학원 2012 국내석사

Nowadays, social network services such as Twitter and Facebook have been widely exploited. Social network services have attracted millions of users, many of whom have combined these services into their daily life. Many people can no longer imagine a life without the social network service. Since it is easy to extend their relations to others in the social network service, users usually have a huge user list. In the user list, various users, not only friends but also unfamiliar users such as celebrities, news media, or/and even unknown users, can be included. A user may add 'friends' with little or no actual connection such as corporations. If people use social network services for a long time, acquaintance lists will be expanded. Then, updated information of intimate friends is covered. In other words, it is difficult to see the information of the familiar friends. Another problem is that users were bored about creating groups for intimate friends from the huge user list. If the friend list is classified into communities, similar users? contents belong to each community. Accordingly, the social network service needs to be divided into the meaningful communities from the friends list of users? for better services. This research aims at the analysis of the huge user list in the social network service and the data mining of the meaningful communities. A mixed method combining Support Vector Machine method with some clustering methods was utilized in order to divide a huge users? list to the meaningful communities. This method needs to be optimized to the social network services. The connection information was collected from 168 seed users. And also from the 17047 following users. The collected data were applied to Support Vector Machine first of all. For Support Vector Machine, the characteristics of the social network service were analyzed to extract ten features. In addition, three machine learning tools were used such as SVMlight, WEKA, and libSVM for feature selection. The experiment of the clustering technique was performed, and the previous Support Vector Machine results were put into graph clustering algorithms in the experiment. The experimental results for detecting communities show that the test result using Support Vector Machine is improved by 10% on average than that without using Support Vector Machine. The first contribution of the research is to detect the meaningful communities from a huge user list in the social network service. The second contribution is to analyze the characteristics of the social network service and to extract many kinds of features for Support Vector Machine.

Support Vector Machine을 활용한 만성질환자의 복약순응 분류 예측

권지혜 경희대학교 공공대학원 2021 국내석사

Purpose: Due to the importance of medication compliance in patients with chronic disease, this study performed secondary data analysis using data from the 2017 Korea Health Panel Study to develop a classification model, using machine learning (ML) and support vector machines (SVMs), for identifying factors that influence and are related to drug compliance. Method: Of the 17,184 household members surveyed, 1,346 adults aged 18 years or older, who were receiving medical care for one or more chronic diseases including hypertension, diabetes, and hyperlipidemia, were included in the analysis. The collected data were analyzed using the STATA 15.1 program, while the predictive classification model for medication compliance was modelled using R version 4.0.4 and SVM, an ML model that is used for both classification and regression. A logistic regression (LR) model was used to compare the performance of the predictive model. Result: The SVM model identified age, number of chronic diseases, hypertension, stress, education level, and sex as factors that influenced medication compliance in patients with chronic disease. Conversely, the LR model identified restrictions to medical services, marital status, hypertension, depression, resolution of questions by doctors, and hyperlipidemia as the influential factors. Among the ML techniques used to construct the predictive model in this study, SVM demonstrated a lower level of performance than LR. The findings of this study indicate the need for identifying a potential for improving the performance of the model through a data normalization procedure by overcoming the limitations of input variable properties. It is also likely that the low specificity may have resulted from an unbalanced data set, leading to poor performance. Conclusion: Different from reported in previous studies, it was evident in this study that mental or physical stress, restrictions to medical services, and resolution of questions by doctors were important factors influencing medication compliance in individuals with hypertension and hyperlipidemia, among other chronic diseases. Such findings indicate the need for additional research to identify yet-to-be-studied factors influencing medication compliance, as well as the establishment of strategies to practice and maintain medication compliance among patients with chronic disease. Furthermore, limitations in data must be overcome, and further research applying various ML techniques must comparatively analyze the suitability and accuracy of predictive classification models, based on the model developed in this study, for medication compliance among patients with chronic disease. Key words patients with chronic disease, medication compliance, machine learning, support vector machine, prediction model 본 연구는 만성질환자의 건강관리에 핵심 요소인 복약 순응에 영향을 미치는 요인을 파악하기 위하여 시도되었다. 본 연구는 머신러닝(Machine learning)기법인 Support Vector Machine 방법을 이용하여 분류 모델을 개발하였으며, 2017년 한국의료패널데이터를 이용한 이차 자료 분석 연구이다. 전체 조사 대상자 17,184명의 가구원 중 고혈압, 당뇨, 고지혈증 중 한가지 이상의 만성질환을 진단받아 의료기관을 이용 중인 18세 이상 성인 가운데 최종 분석에 포함된 대상자 수는 총 1,346명이다. 수집된 자료는 STATA 15.1 프로그램을 이용하여 데이터를 분석하였고, 복약순응 분류 예측 모델은 R Version 4.0.4 프로그램을 사용하여 기계 학습 기법 중 분류와 회귀 모두에 사용되는 Support Vector Machine(SVM) 방법을 이용하여 모델링하였다. 로지스틱 회귀(LR) 모델로 예측 모델의 성능을 비교하였다. 본 연구의 결과 만성질환자의 복약순응에 영향을 미치는 중요한 요인은 SVM 모델에서 연령, 만성질환의 개수, 고혈압, 스트레스, 교육정도, 성별 순으로 나타났다. LR 모델에서는 의료이용 제한, 결혼상태, 고혈압, 우울증, 궁금증 해소, 고지혈증이 중요한 요인으로 나타났다. 본 연구의 예측모델에서는 머신러닝기법 중 SVM 이 LR 보다 저하된 성능을 보였다. 이는 불균형 데이터 집합으로 인한 것으로 본 연구에서는 모델의 특이도가 낮게 도출되어 저조한 성능을 보였다. 추후 입력 변수 속성의 한계점을 보완하고 자료 정규화 방법을 통해 모델의 성능을 높일 수 있는 가능성을 확인할 필요성이 있다. 선행연구의 결과와는 달리, 본 연구에서는 만성질환자의 복약 순응에 영향을 미치는 요인으로 고혈압 또는 고지혈증의 만성질환을 가진 경우, 정신·신체적 스트레스, 의료이용 제한, 의사와의 궁금증 해소가 복약순응의 중요한 요인으로 확인되었다. 이러한 결과를 바탕으로 향후 복약 순응 관련 요인들에 대한 추가적인 연구와 만성질환자의 복약 순응 실천 및 유지 전략 중재 개발이 필요하다. 또한 본 연구의 만성질환자 복약순응 분류 예측 모델을 기초로 향후 데이터의 한계점을 보완하고 지속적인 연구를 통해 다양한 머신러닝기법을 적용한 분류 예측 모델의 적합성 및 정확도를 비교분석해 볼 필요가 있다. 주요어: 만성질환자, 복약순응, 머신러닝, Support Vector Machine, 예측 모델

Intelligent data selection and semi-supervised learning for support vector regression

김동일 서울대학교 대학원 2013 국내박사

Support Vector Regression (SVR), a regression version of Support Vector Machines (SVM), employing Structural Risk Minimization (SRM) principle has become one of the most spotlighted algorithms with the capability of solving nonlinear problems using the kernel trick. Despite of the great generalization performance, there still exist open problems for SVR to overcome. In this dissertation, two major open problems of SVR are studied: (1) training complexity and (2) Semi–Supervised SVR (SS–SVR). Since the training complexity of SVR is highly related to the number of training data n: O(n3), training time complexity and O(n2), the training memory complexity, it makes SVR difficult to be applied to big–sized real–world datasets. In this dissertation, a data selection method, Margin based Data Selection (MDS), was proposed in order to reduce the training complexity. In order to overcome the training complexity problem, reducing the number of training data is an effective approach. Data selection approach is designed to select important or informative data among all training data. For SVR, the most important data are support vectors. By ε–loss foundation and the maximum margin learning, all support vectors of SVR are located on or outside the ε–tube. With multiple sample learning, MDS estimated the margin for all training data, efficiently. MDS selected a subset of data by comparing the margin and ε. Through the experiments conducted on 20 datasets, the performance of MDS was better than the benchmark methods. The training time of SVR including running time of MDS was with 38% ∼ 67% of training time of original datasets. At the same time, the accuracy loss was 0% ∼ 1% of original SVR model. Recently, the size of dataset is getting larger, and data are collected from various applications. Since collecting the labeled data is expensive and time consuming, the fraction of the unlabeled data over the labeled data is getting increased. The conventional supervised learning method uses only labeled data to train. Recently, Semi–Supervised Learning (SSL) has been proposed in order to improve the conventional supervised learning by training the unlabeled data along with the labeled data. In this dissertation, a data generation and selection method for SS–SVR training is proposed. In order to estimate the label distribution of the unlabeled data, Probabilistic Local Reconstruction method (PLR) was employed. In order to get robustness to noisy data, two PLRs (PLRlocal and PLRglobal) were employed and the final label distribution was obtained by the conjugation of 2–PLR. Then, training data were generated from the unlabeled data with their the estimated label distribution. The data generation rate was differed by uncertainty of the labeling. After that, MDS was employed to reduce the training complexity increased by the generated data. Through the experiments conducted on 18 datasets, the proposed method could improve about 10% of the accuracy than the conventional supervised SVR, and the training time of the proposed method including the construction of final SVR was less than 25% of benchmark methods. Two applications are analyzed. For response modeling, SVR based two–stage response modeling, identifying respondents at the first stage and then ranking them according to expected profit at the second stage, was proposed. And MDS was employed in order to reduce the training complexity of two–stage response modeling. The experimental results showed that SVR employed two–stage response model could increase the profit than the conventional response model. MDS reduced the training complexity of SVR to about 60% of original SVR with minimum profit loss. For Virtual Metrology (VM), the proposed SS–SVR method was applied to a real–world VM dataset by using the unlabeled data with the labeled data for training. Data were collected from two pieces of equipment of the photo process. The experimental results showed the proposed SS–SVR method could improve the accuracy about 8% on average than that of the conventional VM model. The accuracy of proposed method was better than benchmark method while the training time of the proposed method was relatively small than benchmark methods.

SUPPORT VECTOR MACHINE을 이용한 생존분석

이보람 인하대학교 대학원 2008 국내석사

Survival analysis is a collection of statistical procedures for data analysis for which the outcome variable of interest is time until an event occurs. Censoring occurs when we have some information about individual survival time, but we don't know the survival time exactly. So we need to adjust the statistical analysis procedures for censored data. A support vector machine (SVM) performs classification by constructing an hyperplane that optimally separates the data into two categories, and separating the set of vector space belonging to two separate classes. In this paper, we study support vector classification with censored data. As a result, rates of miss-classification are quite low. Also the method has better performace for censoring data. 생존분석(Survival Analysis)은 사건의 발생여부에 대한 불확실한 자료(Censored data)가 포함되어 있다는 특징을 가지고 있다. 이런 자료를 분류분석을 하기위해 많은 방법들 중 Support Vector Machine (SVM)방법을 이용한다. Support Vector Machine (SVM)은 간단한 알고리즘을 이용하여 최적의 선형 결정 평면을 찾는 분류방법으로서 직관적인 해석을 제공해 준다. 또한, 벡터공간에서 임의의 경계를 찾아 두 개의 집합을 분류하는 방법으로 주어진 조건하에서 최적의 해를 찾을 수 있다. 본 논문에서는 중도절단된 데이터를 SVM 분류분석을 실시하였다. 그 결과 관측시간에 따른 오분류율이 낮아 모형이 잘 적합함을 알 수 있다.

Incremental class deep support vector data description with auxiliary classifier : 보조적인 분류기를 가지는 점진적인 클래스 심층 서포트 벡터 데이터 디스크립션

백성대 경북대학교 대학원 2018 국내석사

이 연구에서는 심층 학습이 성공할 수 있었던 원인을 분석하고, 모델의 결과를 설명할 수 있도록 기존의 전통적인 기계 학습 중 하나인 Support Vector Data Description을 분석하여, 이 두 가지의 방법의 이점을 얻는 모델인, Deep Support Vector Data Description을 제안한다. 제안하는 모델을 통해 여러 Support Vector Data Description을 심층 학습의 구조로 쌓아 주어진 데이터를 각각 Support Vector Data Description에서 분류할 수 있음을 확인했다. 또한 외부 클래스의 데이터를 잘 못 분류하는 문제를 해소하기 위해 보조적인 분류기로 Support Vector Machine을 Support Vector Data Description에 붙인, Incremental Deep Support Vector Data Description with Auxiliary Classifier 모델을 제안한다. 우리는 제안하는 모델이 목표 클래스의 데이터를 심층적으로 분석하면서 외부 클래스의 데이터를 배제하는 점을 실험을 통해 발견했다. 마지막으로 실험 결과가 제안하는 모델의 유리한 점을 설명한다. In this study, the Support Vector Data Description, which is one of the conventional traditional machine learning methods, is analyzed to measure the cause of the success of deep learning and to explain the result of the model. The model, which obtains the advantages of these two methods, Deep Support Vector Data Description is proposed. Through the proposed model, we confirmed that various Support Vector Data Descriptions can be classified into Support Vector Data Description. In order to solve the problem of misclassifying the data of the external class, we propose the Incremental Deep Support Vector Data Description with Auxiliary Classifier model attached Support Vector Machine as Support Classifier to Support Vector Data Description. We found through experiments that the proposed model analyzed the data of the target class in depth and exclude the data of the external class. Finally, the results of the experiment explained the advantages of the proposed model.

Implementation of Heterogeneous cluster Scheduler for a suitable HEVC Encoder based on Support Vector Machine

반소정 건국대학교 대학원 2018 국내석사

Implementation of Heterogeneous cluster Scheduler for a suitable HEVC Encoder based on Support Vector Machine PAN, SUJING Department of Electronics, Information and Communication Engineering Graduate School of Konkuk University With the continuous improvement of quality of life, the demand for high-quality multimedia terminals is also growing. UHD, 4KTV and other high resolution and high frame rate video terminals are constantly emerging, and 8KTV video terminals will soon also be launched. People's endless demand for audio-visual effects of media brings great challenges to the development of video coding technology. HEVC is a new video compression standard that meets 4KTV, 8KTV and other coding requirements. Compared with the H.264 standard, the compression efficiency is increased by 50% - 70%, and the complexity is increased by 2 - 4 times. Because of the large amount of computation and high memory bandwidth of the HEVC coding algorithm, the design and optimization of HEVC coding and the research and development of the corresponding real time coding system have become the key to the rapid entry of the market application and have become a hot research focus for the market application. Considering the high complexity of the HEVC coding algorithm, it is very effective to develop heterogeneous distributed HEVC video coding by using the universal processor platform to complete the HEVC real-time coding task above the 1080p. Therefore, I propose a cheap heterogeneous distributed HEVC encoder system to replace the high cost encoder platform. The most difficult task in the heterogeneous distributed system is the scheduling task, which makes the whole distributed system in a balanced state, but it often needs to combine the data types of the scheduling task. Therefore, a support vector machine prediction method based on HEVC and embedded platform features is proposed in order to improve the task allocation accuracy of HEVC heterogeneous distributed systems. HEVC 부호기에 적합한 Support Vector Machine 기반 이기종 클러스터 스케줄러 구현 반소정 전자정보통신 학과 건국대학교 일반대학원명 유선 무선 인터넷 속도가 지속적으로 향상됨에 따라 고품질의 멀티미디어 단말기에 대한 수요도 증가하고 있다. 예를 들어UHD, 4KTV 및 기타 고해상도 및 고 프레임 레이트 비디오 단말기가 끊임없이 등장하고 있으며 8KTV 비디오 단말기도 곧 출시될 예정이다. 미디어의 효과에 대한 사람들의 끊임없는 요구는 비디오 코딩 기술의 개발에 커다란 도전 과제를 안겨준다. High Efficiency Video Coding (HEVC )는 4K TV, 8K TV 및 기타 코딩 요구 사항을 충족하는 새로운 비디오 압축 표준이다. H.264 표준과 비교할 경우 압축 효율이 약 50 % ~ 70 % 증가하고 반면 복잡도는 약 2 ~ 4 배가 증가했다. HEVC 코딩 알고리즘의 많은 양의 계산 및 높은 메모리 대역폭으로 인해 HEVC 코딩은 설계 및 최적화와 실시간 코딩 시스템은 연구 및 개발이 중점이 되었다. HEVC 코딩 알고리즘의 높은 복잡성을 고려할 때, 범용 프로세서 플랫폼을 사용하여 1080p 이상의 HEVC 실시간 코딩 작업을 완료함으로써 이기종 분산 HEVC 비디오 코딩을 개발하는 것이 매우 효과적인 것을 알수있다. 따라서 본 논문에서 고가의 인코더 플랫폼을 대체하기 위해 저렴한 가격의 이기종 분산 HEVC 인코더 시스템을 제안하였다. 이기종 분산 시스템에서 가장 어려운 점은 전체 분산 시스템을 균형 있게 만드는 작업이다. 그리기에 본 논문에서는 HEVC 이종 분산 시스템 지연을 줄이고 분산 시스템의 부하 균형을 향상시키기 위해 HEVC 기능과 분산처리 시스템의 플랫폼 퍼포먼스 기반으로 support vector machine 기계학습 방법을 제안하고 구현을 하였다. 결과적으로 전체 시스템은 다른 이기종 분산처리를 한 HEVC encoder에 비해 화면 소실이 없는 하에서 약 1.45정도의 성능 향상을 얻었다.