Learning with continuous piecewise linear decision trees
作者：
Highlights：
• (G)PWL-DT is proposed as a promising alternative of PWC-DT in tree learning methods.
• Indicator functions are replaced by ReLUs to do domain partitions for the leaf nodes.
• For each leaf, a continuous PWL decision rule is constructed based on nested ReLUs.
• (G)PWL-DT overcomes discontinuity of typical DTs and introduces piecewise linearity.
• Given the same DT structure, (G)PWL-DT can be more flexible than PWC-DT in boundaries.
摘要
•(G)PWL-DT is proposed as a promising alternative of PWC-DT in tree learning methods.•Indicator functions are replaced by ReLUs to do domain partitions for the leaf nodes.•For each leaf, a continuous PWL decision rule is constructed based on nested ReLUs.•(G)PWL-DT overcomes discontinuity of typical DTs and introduces piecewise linearity.•Given the same DT structure, (G)PWL-DT can be more flexible than PWC-DT in boundaries.
论文关键词：Decision tree,Continuity,Piecewise linearity,Domain partition
论文评审过程：Received 6 May 2020, Revised 4 October 2020, Accepted 1 November 2020, Available online 3 November 2020, Version of Record 24 January 2021.
论文官网地址：https://doi.org/10.1016/j.eswa.2020.114214
