基於區域確定性的群集設計主動學習樣本資料挑選的方法

Efficient Active Learning Based on Localized Uncertainty Clusters

吳婉萍(Wu,WangPing)

指導教授 ： 李新林

國立中正大學/工學院/資訊工程學系/碩士(2012年)

摘要

無資料

關鍵字

分群

並列摘要

There are many approaches which incorporate clustering into active learning
for avoiding selecting the similar data points in the active learning. Yet,
the traditional clustering methods do not consider increasing the accuracy
of the active learning. There are two main purposes in this thesis. One is
that more uncertain representatives of clusters are generated in clustering
to increase the accuracy of the classification, and the other is that clustering
with an unknown number of clusters. In our method, data points with
similar in the local uncertainty, the small difference of coordinates and the
large overlapped neighborhood will be collected into the same cluster. And
the idea of certainty-based active learning(CBAL), which a local classifier is
built by using neighbors, is used for finding the appropriate size of neighborhood.
In our approach, the generated representatives can represent the effect
of all data points in the same cluster to the classifier. Moreover, more uncertain
representatives are generated, which will increase the accuracy of active
learning. In addition, we propose a new clustering method which uses the
value of local distance-based outlier factor(LDOF) to expand size of clusters
and uses the distance measurement metric based on local uncertainty and
overlapped neighborhood (LNC formula) to measure the similarity between
data points. Finally, the experimental results show that the proposed method
can select more uncertain training data in the synthetic dataset. And, in the
UCI datasets , the accuracy and running time are also better.

並列關鍵字

clustering

參考文獻

Advances in Neural Information Processing Systems (NIPS), volume 15,

pp. 561-568. MIT Press, 2003.

with Improved Label Complexity”, appears in: arXiv:1108.1766v1,

for Error-Reduction Sampling in Active Learning”, appears in: Data

[4] Viet-Vu Vu, “ Active Learning for Semi-Supervised K-Means Clustering”,