TY - JOUR
T1 - Robust expected model change for active learning in regression
AU - Park, Sung Ho
AU - Kim, Seoung Bum
N1 - Funding Information:
This work was supported by Brain Korea PLUS, the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Science, ICT, and Future Planning (NRF-2016R1A2B1008994) and the Ministry of Trade, Industry & Energy under the Industrial Technology Innovation Program (R1623371).
PY - 2020/2/1
Y1 - 2020/2/1
N2 - Active learning methods have been introduced to reduce the expense of acquiring labeled data. To solve regression problems with active learning, several expected model change maximization strategies have been developed to select the samples that are likely to greatly affect the current model. However, some of the selected samples may be outliers, which can result in poor estimation performance. To address this limitation, this study proposes an active learning framework that adopts an expected model change that is robust for both linear and nonlinear regression problems. By embedding local outlier probability, the learning framework aims to avoid outliers when selecting the samples that result in the greatest change to the current model. Experiments are conducted on synthetic and benchmark data to compare the performance of the proposed method with that of existing methods. The experimental results demonstrate that the proposed active learning algorithm outperforms its counterparts.
AB - Active learning methods have been introduced to reduce the expense of acquiring labeled data. To solve regression problems with active learning, several expected model change maximization strategies have been developed to select the samples that are likely to greatly affect the current model. However, some of the selected samples may be outliers, which can result in poor estimation performance. To address this limitation, this study proposes an active learning framework that adopts an expected model change that is robust for both linear and nonlinear regression problems. By embedding local outlier probability, the learning framework aims to avoid outliers when selecting the samples that result in the greatest change to the current model. Experiments are conducted on synthetic and benchmark data to compare the performance of the proposed method with that of existing methods. The experimental results demonstrate that the proposed active learning algorithm outperforms its counterparts.
KW - Active learning
KW - Expected model change maximization
KW - Local outlier probability
KW - Stochastic gradient descent
UR - http://www.scopus.com/inward/record.url?scp=85069746064&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85069746064&partnerID=8YFLogxK
U2 - 10.1007/s10489-019-01519-z
DO - 10.1007/s10489-019-01519-z
M3 - Article
AN - SCOPUS:85069746064
VL - 50
SP - 296
EP - 313
JO - Applied Intelligence
JF - Applied Intelligence
SN - 0924-669X
IS - 2
ER -