林智仁臺灣大學：資訊工程學研究所范榮恩Fan, Rong-EnRong-EnFan2007-11-262018-07-052007-11-262018-07-052007http://ntur.lib.ntu.edu.tw//handle/246246/54137多標籤分類近年來在各種應用中越來越普遍，比如在文件分類或多媒體搜尋系統。為滿足不同應用的需求，許多評分標準被提出。目前最常被用來解決多標籤分類的方法為雙類比對。此方法替每個標籤創造一個判斷函數。對於某些應用而言，調整判斷函數的門檻值會增進效能。在本篇論文中，我們針對門檻值的選擇進行深入探討。並透過真實應用產生的資料來展示這類方法的有用之處。Multi-label classification becomes more and more popular in recent years. It is used in, for example, text categorization or multimedia retrieval systems. Many evaluation criteria are proposed for different application needs. A commonly used approach for multi-label classification is the binary method, which constructs a decision function per label. For some applications, adjusting thresholds in decision functions improves the performance. This thesis gives a comprehensive study on the selection of thresholds. Experiments on several real-world data sets demonstrate the usefulness of some simple selection strategies.口試委員審定書 i 中文摘要 ii ABSTRACT iii LIST OF TABLES vi CHAPTER I. Introduction 1 II. Binary Method and Evaluation Measures 4 2.1 The Binary Method 4 2.2 Evaluation Criteria 5 2.2.1 Exact Match Ratio 6 2.2.2 Macro-average and Micro-average F-measure 6 2.2.3 Ranking Based Measures 7 2.3 Issues on Optimizing Different Measures 9 III. Optimize Measures via Supervised Threshold Setting 14 3.1 Supervised Threshold Setting in Binary Method 14 3.1.1 The SVM.1-type Methods 16 3.2 Real-World Data Sets 21 3.2.1 Yahoo! 22 3.2.2 scene 23 3.2.3 yeast 24 3.2.4 OHSUMED 25 3.2.5 RCV1-V2 25 3.3 Experiments 25 3.3.1 Experimental Settings 26 3.3.2 Optimizing Macro-average F-measure 27 3.3.3 Optimizing Micro-average F-measure 30 3.3.4 Optimizing Exact Match Ratio 32 3.3.5 Discussion and Conclusion 35 IV. Conclusions 38 BIBLIOGRAPHY 39458404 bytesapplication/pdfen-US多標籤分類評分標準門檻值選擇雙類比對支撐向量機multi-label classificationevaluation criteriasupervised threshold settingbinary methodsupport vector machinesthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/54137/1/ntu-96-R94922033-1.pdf