Streaming Decision Trees for Lifelong Learning

• Published in: Machine Learning and Knowledge Discovery in Databases. Research Track Vol. 12975; pp. 502 - 518
• Main Authors: Korycki, Łukasz, Krawczyk, Bartosz
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Catastrophic forgetting; Continual learning; Data streams; Decision trees; Lifelong learning
• ISBN: 3030864855; 9783030864859
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-86486-6_31

Lifelong learning models should be able to efficiently aggregate knowledge over a long-term time horizon. Comprehensive studies focused on incremental neural networks have shown that these models tend to struggle with remembering previously learned patterns. This issue known as catastrophic forgetting has been widely studied and addressed by several different approaches. At the same time, almost no research has been conducted on online decision trees in the same setting. In this work, we identify the problem by showing that streaming decision trees (i.e., Hoeffding Trees) fail at providing reliable long-term learning in class-incremental scenarios, which can be further generalized to learning under temporal imbalance. By proposing a streaming class-conditional attribute estimation, we attempt to solve this vital problem at its root, which, ironically, lies in leaves. Through a detailed experimental study we show that, in the given scenario, even a rough estimate based on previous conditional statistics and current class priors can significantly improve the performance of streaming decision trees, preventing them from catastrophically forgetting earlier concepts, which do not appear for a long time or even ever again.