輔助學習式多標籤主動學習演算法

Abstract

在多標籤分類的應用中，由於標籤費用高昂的關係，使得多標籤主動學習開始成為一個熱門的研究領域，而其中一個最新被提出來的演算法 MMC (maximum loss reduction with maximum confidence)，無論是在學習還是詢問的步驟中，都高度依賴著一個特定的多標籤分類器：一對多支持向量機 (binary relevance support vector machine) ，但是MMC的這種高度依賴性並不明確是否是必要的，在本論文中，我們提出了一個一般性的多標籤主動學習框架，這個框架移除了MMC的高度依賴性，並且擁有可參數化的三個元件：用來做決策的主要學習者；用來輔助詢問的次要學習者；以及一個詢問的策略。MMC以及許多其他的演算法都可以被視為這個框架的特例。而基於這個框架，我們提出了兩個不同於以往的詢問策略，HLR (Hamming loss reduction) 和SHLR (soft Hamming loss reduction)，並在許多不同的主要/次要學習者的組合上測試了這些詢問的策略的好壞。在許多的實驗中都顯示，我們所提出的SHLR，無論是在何種衡量基準以及主要/次要學習者的組合上都擁有最穩定良好的表現。

Multi-label active learning is an important problem because of the expensive labeling cost in multi-label classification applications. A state-of-the-art approach for multi-label active learning, maximum loss reduction with maximum confidence (MMC), heavily depends on the binary relevance support vector machine in both learning and querying. Nevertheless, it is not clear whether the heavy dependence is necessary or unrivaled. In this work, we extend MMC to a more general framework that removes the heavy dependence and clarifies the roles of each component in MMC. In particular, the framework is characterized by a major learner for making predictions, an auxiliary learner for helping with query decisions and a query criterion based on the disagreement between the two learners. The framework takes MMC and several baseline multi-label active learning algorithms as special cases. With the flexibility of the general framework, we design two criteria other than the one used by MMC. We also explore the possibility of using learners other than the binary relevance support vector machine for multi-label active learning. Experimental results demonstrate that a new criterion, soft Hamming loss reduction, is usually better than the original MMC criterion across different pairs of major/auxiliary learners, and validate the usefulness of the proposed framework.