没有合适的资源?快使用搜索试试~ 我知道了~
2019-KDD-Automating Feature Subspace Exploration via Multi-Agent
需积分: 0 0 下载量 179 浏览量
2022-08-04
14:54:03
上传
评论
收藏 1.6MB PDF 举报
温馨提示
试读
9页
2019-KDD-Automating Feature Subspace Exploration via Multi-Agent Reinforcement Learning-图表示+多智能体+以奖励改善特征之协调-rrrr1
资源详情
资源评论
资源推荐
Automating Feature Subspace Exploration via Multi-Agent
Reinforcement Learning
Kunpeng Liu
University of Central Florida
Florida, US
kunpengliu@knights.ucf.edu
Yanjie Fu
∗
University of Central Florida
Florida, US
Yanjie.Fu@ucf.edu
Pengfei Wang
CNIC, Chinese Academy of Sciences
Beijing, China
wpf@cnic.cn
Le Wu
Hefei University of Technology
Hefei, China
lewu@hfut.edu.cn
Rui Bo
Missouri Univ. of Sci. and Tech.
Missouri, US
rbo@mst.edu
Xiaolin Li
Nanjing University
Nanjing, China
lixl@nju.edu.cn
ABSTRACT
Feature selection is the preprocessing step in machine learning
which tries to select the most relevant features for the subsequent
prediction task. Eective feature selection could help reduce dimen-
sionality, improve prediction accuracy and increase result compre-
hensibility. It is very challenging to nd the optimal feature subset
from the subset space as the space could be very large. While much
eort has been made by existing studies, reinforcement learning
can provide a new perspective for the searching strategy in a more
global way. In this paper, we propose a multi-agent reinforcement
learning framework for the feature selection problem. Specically,
we rst reformulate feature selection with a reinforcement learning
framework by regarding each feature as an agent. Then, we obtain
the state of environment in three ways, i.e., statistic description,
autoencoder and graph convolutional network (GCN), in order to
make the algorithm better understand the learning progress. We
show how to learn the state representation in a graph-based way,
which could tackle the case when not only the edges, but also the
nodes are changing step by step. In addition, we study how the co-
ordination between dierent features would be improved by more
reasonable reward scheme. The proposed method could search the
feature subset space globally and could be easily adapted to the
real-time case (real-time feature selection) due to the nature of
reinforcement learning. Also, we provide an ecient strategy to
accelerate the convergence of multi-agent reinforcement learning.
Finally, extensive experimental results show the signicant im-
provement of the proposed method over conventional approaches.
CCS CONCEPTS
• Computing methodologies → Multi-agent reinforcement
learning; Feature selection.
∗
Contact Author.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
KDD ’19, August 4–8, 2019, Anchorage, AK, USA
© 2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-6201-6/19/08.. . $15.00
https://doi.org/10.1145/3292500.3330868
KEYWORDS
feature selection; automated exploration; multi-agent reinforce-
ment learning
ACM Reference Format:
Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li. 2019.
Automating Feature Subspace Exploration via Multi-Agent Reinforcement
Learning. In The 25th ACM SIGKDD Conference on Knowledge Discovery and
Data Mining (KDD ’19), August 4–8, 2019, Anchorage, AK, USA. ACM, New
York, NY, USA, 9 pages. https://doi.org/10.1145/3292500.3330868
1 INTRODUCTION
Feature selection aims to select an optimal subset of relevant fea-
tures for a downstream predictive task [
2
,
34
]. Eective feature
selection can help to reduce dimensionality, shorten training times,
enhance generalization, avoid overtting, improve predictive ac-
curacy, and provide better interpretation and explanation. In this
paper, we study the problem of automated feature subspace explo-
ration for improving downstream predictive tasks.
Prior studies in feature selection can be grouped into three cate-
gories: (i) lter methods (e.g., univariate feature selection [
5
,
33
],
correlation based feature selection [
10
,
34
]), in which features are
ranked by a specic score; (ii) wrapper methods (e.g., evolution-
ary algorithms [
11
,
31
], branch and bound algorithms [
13
,
20
]), in
which optimal feature subset is identied by a search strategy that
collaborates with predictive tasks; (iii) embedded methods (e.g.,
LASSO [
29
], decision tree [
26
]), in which feature selection is part
of the optimization objective of predictive tasks. However, these
studies have shown not just strengths but also some limitations. For
example, lter methods ignore the feature dependencies and inter-
actions between feature selection and predictors. Wrapper methods
have to search a very large feature space of 2
N
feature subspace
candidates, where
N
is the feature number. Embedded methods
are subject to the strong structured assumptions of predictive mod-
els. As can be seen, feature selection is a complicated process that
requires (i) strategic design of feature signicance measurement,
(ii) accelerated search of near-optimized feature subset, and (iii)
meaningful integration of predictive models.
Reinforcement learning can interact with environments, learn
from action rewards, balance exploitation and exploration, and
search for long-term optimal decisions [
17
,
35
]. These traits pro-
vide great potential to automate feature subspace exploration. Ex-
isting studies [
4
,
14
] create a single agent to make decisions. In
Research Track Paper
KDD ’19, August 4–8, 2019, Anchorage, AK, USA
207
VashtaNerada
- 粉丝: 18
- 资源: 297
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
评论0