Inductive algorithms rely strongly on their representational biases. Representational inadequacy can be mitigated by constructive induction. This paper introduces the notion of relative gain measure and describes a new constructive induction algorithm (GALA) which generates a small number of new attributes from existing nominal or real-valued attributes. Unlike most previous research on constructive induction, our techniques are designed for use in preprocessing data set for subsequent use by any standard selective learning algorithms. We present results which demonstrate the effectiveness of GALA on both artificial and real domains with respect to C4.5 and CN2.
|主出版物標題||Advances in Artificial Intelligence - 11th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, AI 1996, Proceedings|
|出版狀態||Published - 1 一月 1996|
|事件||11th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, AI 1996 - Toronto, Canada|
持續時間: 21 五月 1996 → 24 五月 1996
|名字||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Conference||11th Biennial Conference of the Canadian Society for Computational Studies of Intelligence, AI 1996|
|期間||21/05/96 → 24/05/96|