TabNet: Attentive Interpretable Tabular Learning

읽은 날	2020.09.15	학술지	arXiv preprint arXiv
제목	TabNet: Attentive Interpretable Tabular Learning
저자	Sercan ¨ O. Arık, Tomas Pfister
한줄요약	Tabular data를 처리하는 새로운 neural network인 TabNet의 이점.
초록	We propose a novel high-performance and interpretable canonical deep tabular data learning architecture, TabNet. TabNet uses sequential attention to choose which features to reason from at each decision step, enabling interpretability and more efficient learning as the learning capacity is used for the most salient features. We demonstrate that TabNet outperforms other neural network and decision tree variants on a wide range of non-performance-saturated tabular datasets and yields interpretable feature attributions plus insights into the global model behavior. Finally, for the first time to our knowledge, we demonstrate self-supervised learning for tabular data, significantly improving performance with unsupervised representation learning when unlabeled data is abundant.
키워드	TabNet, tabular data, deep neural network, deep learning
의의	정확도 높은 tabular data learning algorithm 개발
비판점