An Active Learning Algorithm for Semi-Supervised Clustering and a Semi-Supervised Active Learning Framework for Distance Metric Learning Based on Locally Propagated Linear Reconstruction

View Statistics Email Alert RSS Feed

Information

Details

Project title

Code/計畫編號

NSC102-2221-E019-058

Translated Name/計畫中文名

基於區域傳遞式線性重建技巧的一個適用於半監督式分群的主動式學習演算法與一個半監督、主動式距離學習架構

Project Coordinator/計畫主持人

Chin-Chun Chang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=3098747

Year

2013

Start date/計畫起

01-08-2013

Expected Completion/計畫迄

31-07-2014

Bugetid/研究經費

538千元

ResearchField/研究領域

資訊科學--軟體

"主動式學習可為接下來的學習工作詢問最具資訊力的樣本的類別標籤，藉此可減少標示所有訓練樣本的類別標籤所花費的代價。在這個兩年計劃裡，我們將研究適合半監督式分群的主動式學習演算法與一個適合距離學習之半監督、主動式學習框架。尤其，我們將發展一個稱為區域傳遞式線性重建的技巧，來產生代表樣本集合。在代表樣本集合裡的樣本不但對區域線性重建是重要的，也對附近樣本重建具有重大的影響力。接下來，配合對接續的學習工作合適的樣本挑選準則，我們可以從代表樣本集合裡挑出有效的詢問樣本。這兩年的研究主題如下:  第一年、基於區域傳遞式線性重建技巧的一個適用於半監督式分群的主動式學習演算法:半監督式分群具有使用少量監督資訊來產生比傳統分群法更精確的分群結果的能力。在半監督式分群裡，監督資訊可以用來做資料層級與空間層級的推論，並且最好能捕捉到資料的群結構。目前針對半監督式分群的主動式學習法裡，幾乎沒有一個方法可以同時考慮這三種監督資訊的使用方式來產生詢問樣本。在第一年，我們將發展這種主動式學習法。  第二年、基於區域傳遞式線性重建的一個半監督、主動式距離學習架構:距離學習主要是學習一個距離函數來定義樣本間恰當的距離關係。在第二年，我們將研究一個半監督、主動式距離學習架構。首先，我們將從代表樣本集合裡挑出有效的詢問樣本。然後，使用標籤傳遞方式對來標示未標籤資料。最後，使用傳統的距離學習演算法與所有訓練資料來學習得到距離函數。區域傳遞式線性重建是這兩年計畫的核心技術，這個技術對相關應用也深具價值。非常值得研究與發展相關應用。""Active learning queries the label of the most informative sample for subsequent learning tasks. In this two-year project, we shall consider the problem of active learning for semi-supervised clustering and design a semi-supervised active learning framework for distance metric learning. In particular, we shall develop a new technique for active learning which is called locally propagated linear reconstruction. This technique will be used to produce a representative-sample set which includes samples being important to local reconstruction, and having strong influence on the local reconstruction of nearby samples. With the criterion suitable for the subsequent learning task, effective query samples may be selected from the representative-sample set. The topics of this two-year project are as follows.  The first year: An active learning algorithm for semi-supervised clustering based on locally propagated linear reconstruction. Semi-supervised clustering is data clustering with a small amount of supervised information, and may improve the accuracy of traditional data clustering. In semi-supervised clustering, the supervised information may be used for instance-level and spatial-level implications, and is also desired to reflect underlying cluster structures. Few existing algorithms of active learning for semi-supervised clustering produce queries by taking account of these three usages of supervised information simultaneously. In the first year, we shall aim at such an active-learning algorithm.  The second year: A semi-supervised active learning framework for distance metric learning based on locally propagated linear reconstruction. Distance metric learning aims at a distance function defining proper distance relationships among samples. In the second year, we shall aim at a semi-supervised active learning framework for distance metric learning. First, the query sample will be selected from the representative-sample set. After acquiring the label of the query sample, unlabeled samples are labeled through label propagation. At last, a traditional distance metric learning will be applied to learn the distance metric. Locally propagated linear reconstruction is the core technology in this two-year project, and valuable to related applications. It is very worthy of investigating this technology and its applications. "

Keyword(s)

主動式學習
區域線性重建
半監督式分群
半監督式距離學習
Active learning
Locally linear reconstruction
Semi-supervised clustering
Semi-supervised distance metric learning

DSpace CRIS

An Active Learning Algorithm for Semi-Supervised Clustering and a Semi-Supervised Active Learning Framework for Distance Metric Learning Based on Locally Propagated Linear Reconstruction

Details

Description