【摘要】：多標(biāo)記學(xué)習(xí)(multi-label learning,MLL)任務(wù)處理一個(gè)示例對(duì)應(yīng)多個(gè)標(biāo)記的情況,其目標(biāo)是學(xué)習(xí)一個(gè)從示例到相關(guān)標(biāo)記集合的映射.在MLL中,現(xiàn)有方法一般都是采用均勻標(biāo)記分布假設(shè),也就是各個(gè)相關(guān)標(biāo)記(正標(biāo)記)對(duì)于示例的重要程度都被當(dāng)作是相等的.然而,對(duì)于許多真實(shí)世界中的學(xué)習(xí)問(wèn)題,不同相關(guān)標(biāo)記的重要程度往往是不同的.為此,標(biāo)記分布學(xué)習(xí)將不同標(biāo)記的重要程度用標(biāo)記分布來(lái)刻畫,已經(jīng)取得很好的效果.但是很多數(shù)據(jù)中卻僅包含簡(jiǎn)單的邏輯標(biāo)記而非標(biāo)記分布.為解決這一問(wèn)題,可以通過(guò)挖掘訓(xùn)練樣本中蘊(yùn)含的標(biāo)記重要性差異信息,將邏輯標(biāo)記轉(zhuǎn)化為標(biāo)記分布,進(jìn)而通過(guò)標(biāo)記分布學(xué)習(xí)有效地提升預(yù)測(cè)精度.上述將原始邏輯標(biāo)記提升為標(biāo)記分布的過(guò)程,定義為面向標(biāo)記分布學(xué)習(xí)的標(biāo)記增強(qiáng).首次提出了標(biāo)記增強(qiáng)這一概念,給出了標(biāo)記增強(qiáng)的形式化定義,總結(jié)了現(xiàn)有的可以用于標(biāo)記增強(qiáng)的算法,并進(jìn)行了對(duì)比實(shí)驗(yàn).實(shí)驗(yàn)結(jié)果表明:使用標(biāo)記增強(qiáng)能夠挖掘出數(shù)據(jù)中隱含的標(biāo)記重要性差異信息,并有效地提升MLL的效果.
[Abstract]:The multi-tag learning (multi-label learning,MLL) task deals with a case where an example corresponds to multiple tags, the goal of which is to learn a mapping from an example to a collection of related tags. In MLL, the existing methods generally adopt the assumption of uniform label distribution, that is, the importance of each relevant marker (positive marker) to the example is considered to be equal. However, for many real-world learning problems, the importance of different related markers is often different. For this reason, marker distribution learning describes the importance of different markers by label distribution, and has achieved good results. But a lot of data contains only simple logical tags, not tag distributions. In order to solve this problem, we can transform logical markers into tag distribution by mining the difference information of marker importance contained in training samples, and then effectively improve the prediction accuracy through label distribution learning. The above process of upgrading the original logical tag to the label distribution is defined as the tag enhancement oriented to the label distribution learning. The concept of tag enhancement is proposed for the first time, the formal definition of tag enhancement is given, and the existing algorithms that can be used for tag enhancement are summarized and compared with each other. The experimental results show that the use of marker enhancement can mine the hidden information of significance difference and improve the effect of MLL effectively.
【作者單位】： 東南大學(xué)計(jì)算機(jī)科學(xué)與工程學(xué)院;計(jì)算機(jī)網(wǎng)絡(luò)和信息集成教育部重點(diǎn)實(shí)驗(yàn)室(東南大學(xué));軟件新技術(shù)與產(chǎn)業(yè)化協(xié)同創(chuàng)新中心(南京大學(xué));無(wú)線通信技術(shù)協(xié)同創(chuàng)新中心(東南大學(xué));
【基金】：國(guó)家自然科學(xué)基金優(yōu)秀青年科學(xué)基金項(xiàng)目(61622203) 江蘇省自然科學(xué)基金杰出青年基金項(xiàng)目(BK20140022)~~
【分類號(hào)】：TP301.6
，

本文編號(hào)：2265495