Safe incomplete label distribution learning
作者：
Highlights：
• We have proposed a novel method approach, i.e., SILDL, for safe label distribution learning.
• We have used the squared loss in objective function which could be formulated as a simple convex quadratic program and could be easily solved.
• We have defined an equivalent formulation to find out the best result from multiple baseline methods, and we have integrated multiple incomplete supervised learners to maximize the worst-case performance gain against the best baseline method.
• We have proved that the SILDL approach is provably safe with mild conditions.
摘要
•We have proposed a novel method approach, i.e., SILDL, for safe label distribution learning.•We have used the squared loss in objective function which could be formulated as a simple convex quadratic program and could be easily solved.•We have defined an equivalent formulation to find out the best result from multiple baseline methods, and we have integrated multiple incomplete supervised learners to maximize the worst-case performance gain against the best baseline method.•We have proved that the SILDL approach is provably safe with mild conditions.
论文关键词：Label distribution learning,Safeness,Incomplete supervised learning
论文评审过程：Received 4 May 2021, Revised 29 November 2021, Accepted 29 December 2021, Available online 31 December 2021, Version of Record 6 January 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108518
