A Deep Reinforcement Learning Based Feature Selector
A Deep Reinforcement Learning Based Feature Selector
复制标题
DOI：
10.1007/978-981-16-0010-4_33
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Yiran Cheng;Kazuhiko Komatsu;Masayuki Sato;Hiroaki Kobayashi
中科院分区：
文献类型：
--
作者：
Yiran Cheng;Kazuhiko Komatsu;Masayuki Sato;Hiroaki Kobayashi
在数据挖掘和机器学习领域，对高维数据进行分析和分类一直是研究人员和工程技术人员面临的挑战。为了最大限度地减少分类误差，识别和选择最具特征性的特征以及从高维数据中去除不相关的特征是至关重要的。特征选择算法提供了从所有候选特征集中找到紧凑的有价值的特征子集的有效过程。在过去的研究中，大多数特征选择方法将这类问题视为一个评分问题，每个特征被单独评估，并选择排名靠前的特征。在该研究中，将向特征子集添加新特征视为马尔可夫决策过程，并将特征选择任务形式化为强化学习问题。提出了一种新的基于深度强化学习的特征选择器(DRLFS)，并结合一种动态随机性策略和两种搜索协议来解决探索与开发的两难问题，以逐步逼近最优子集。在不同基准数据集上的实验结果证明了本文提出的方法具有良好的特征选择能力。
In the field of data mining and machine learning, it is a challenge for researchers and engineers to analyze and classify the high-dimensional data. In order to minimize the classification error, it is critical to identify and select the most characterizing features as well as remove the irrelevant features from the high-dimensional data. Feature selection algorithms provide effective processes to find a compact valuable feature subset from the set of all the candidate features. In past researches, most feature selection methods treat such a problem as a scoring problem, where each feature is evaluated individually and top-ranked features are selected. In this study, adding new features into the feature subset is considered as a Markov Decision Process, and the Feature Selection task is formalized as a reinforcement learning problem. This paper proposes a novel Deep Reinforcement Learning based Feature Selector (DRLFS) along with a dynamic randomness policy and two search protocols to tackle the exploration versus exploitation dilemma for a gradual approach to the optimum subset. The experimental results on various benchmark datasets prove the promising feature selection ability of the proposal of this paper.