奥田 遼, 吉川 友也, “決定木と深層ニューラルネットを用いた解釈可能な予測モデル,” 第47回情報論的学習理論と機械学習研究会, 神奈川, Sep. 2022.
本研究では Decision Tree の解釈性と，Deep Neural Networks (DNN) の予測精度を両立する手法を提案する. Random Forest は特徴空間を複数の statement が示す境界によって分割する手法として，解釈性に優れている.ただし，境界の数が多すぎると解釈が難しいため，モデルの近似を用いて予測精度を保ちつつ境界の数を減らす研究が存在する.これらの手法は境界の数を減らせるものの， 新たに得られたモデルの予測精度が近似前 のモデルの予測精度を下回りやすい. そこで本研究では各サンプルについて少数の statement を選択し，それらに付与された重みを用いて予測を行う DNN を提案する. 実験においては，Random Forest 単体よりも精度が向上したことを示す.
In this study, we propose a method that achieves both interpretability of Decision Tree and the prediction accuracy of Deep Neural Networks (DNN). Random Forest is a method that divides the feature space by the boundaries indicated by multiple statements and has good interpretability. However, interpretation becomes difficult, if the number of boundaries is too large. There are existing methods that use an approximated model to reduce the number of boundaries while preserving the prediction accuracy.
Although these methods can reduce the number of boundaries, the prediction accuracy of the newly obtained model tends to be lower than that of the model before approximation. Therefore, in this study, we propose a DNN that selects a small number of statements for each sample and makes predictions using the weights assigned to these statements. In experiments, we show that the proposed method improves the accuracy more than Random Forest alone.