[인공지능학술대회] CaRFE: Candidates Recursive Feature Elimination for Black-Box Model - 송영재, 김광수
- 인공지능 융합 연구실
- 조회수236
- 2021-10-05
Abstract
RFE(Recursive Feature Elimination) is a widely-used greedy algorithm for a feature selection. However, the greedy feature selection does not guarantee obtaining an optimal feature subset as there is no method to predict perfectly precise feature importance for a black-box model such as a deep neural network. Thus, this paper proposes a novel variation of RFE for the black-box model, namely CaRFE(Candidates RFE). Through two experiments in the solar power forecasting dataset, we empirically show that CaRFE outperforms RFE in model performance and can reach a nearly optimal feature subset.