没有合适的资源?快使用搜索试试~ 我知道了~
A Review of Unsupervised Feature Learning and Deep Learning for Time-Series Modeling
资源推荐
资源详情
资源评论
http://www.diva-portal.org
This is version of a paper published in Pattern Recognition Letters.
Citation for the original published paper (version of record):
Längkvist, M., Karlsson, L., Loutfi, A. (2014)
A review of unsupervised feature learning and deep learning for time-series modeling.
Pattern Recognition Letters, 42(1): 11-24
http://dx.doi.org/10.1016/j.patrec.2014.01.008
Access to the published version may require subscription.
N.B. When citing this work, cite the original published paper.
Permanent link to this version:
http://urn.kb.se/resolve?urn=urn:nbn:se:oru:diva-34597
A Review of Unsupervised Feature Learning and Deep
Learning for Time-Series Modeling
Martin Längkvist
a,
∗
, Lars Karlsson
a
, Amy Lout
a
a
Applied Autonomous Sensor Systems, School of Science and Technology, Örebro
University, SE-701 82, Örebro, Sweden
Abstract
This paper gives a review of the recent developments in deep learning and un-
supervised feature learning for time-series problems. While these techniques
have shown promise for modeling static data, such as computer vision, ap-
plying them to time-series data is gaining increasing attention. This paper
overviews the particular challenges present in time-series data and provides a
review of the works that have either applied time-series data to unsupervised
feature learning algorithms or alternatively have contributed to modications
of feature learning algorithms to take into account the challenges present in
time-series data.
Keywords:
time-series, unsupervised feature learning, deep learning
1. Introduction and Background
Time is a natural element that is always present when the human brain
is learning tasks like language, vision and motion. Most real-world data
has a temporal component, whether it is measurements of natural processes
∗
Corresponding author
Email addresses:
martin.langkvist@oru.se
(Martin Längkvist),
lars.karlsson@oru.se
(Lars Karlsson),
amy.loutfi@oru.se
(Amy Lout)
Preprint submitted to Pattern Recognition Letters March 24, 2014
(weather, sound waves) or man-made (stock market, robotics). Analysis of
time-series data has been the subject of active research for decades (Keogh
and Kasetty, 2002; Dietterich, 2002) and is considered by Yang and Wu
(2006) as one of the top 10 challenging problems in data mining due to
its unique properties. Traditional approaches for modeling sequential data
include the estimation of parameters from an assumed time-series model,
such as autoregressive models (Lütkepohl, 2005) and Linear Dynamical Sys-
tems (LDS) (Luenberger, 1979), and the popular Hidden Markov Model
(HMM) (Rabiner and Juang, 1986). The estimated parameters can then
be used as features in a classier to perform classication. However, more
complex, high-dimensional, and noisy real-world time-series data cannot be
described with analytical equations with parameters to solve since the dy-
namics are either too complex or unknown (Taylor, 2009) and traditional
shallow methods, which contain only a small number of non-linear opera-
tions, do not have the capacity to accurately model such complex data.
In order to better model complex real-world data, one approach is to
develop robust features that capture the relevant information. However, de-
veloping domain-specic features for each task is expensive, time-consuming,
and requires expertise of the data. The alternative is to use unsupervised
feature learning (Bengio and LeCun, 2007; Bengio et al., 2012; Erhan et al.,
2010) in order to learn a layer of feature representations from unlabeled data.
This has the advantage that the unlabeled data, which is plentiful and easy
to obtain, is utilized and that the features are learned from the data instead
of being hand-crafted. Another benet is that these layers of feature repre-
sentations can be stacked to create deep networks, which are more capable
2
of modeling complex structures in the data. Deep networks have been used
to achieve state-of-the-art results on a number of benchmark data sets and
for solving dicult AI tasks. However, much focus in the feature learning
community has been on developing models for static data and not so much
on time-series data.
In this paper we review the variety of feature learning algorithms that
has been developed to explicitly capture temporal relationships as well as the
various time-series problems that they have been used on. The properties of
time-series data will be discussed in Section 2 followed by an introduction to
unsupervised feature learning and deep learning in Section 3. An overview
of some common time-series problems and previous work using deep learning
is given in Section 4. Finally, conclusions are given in Section 5.
2. Properties of time-series data
Time-series data consists of sampled data points taken from a continuous,
real-valued process over time. There are a number of characteristics of time-
series data that make it dierent from other types of data.
Firstly, the sampled time-series data often contain much noise and have
high dimensionality. To deal with this, signal processing techniques such
as dimensionality reduction techniques, wavelet analysis or ltering can be
applied to remove some of the noise and reduce the dimensionality. The use
of feature extraction has a number of advantages (Nanopoulos et al., 2001).
However, valuable information could be lost and the choice of features and
signal processing techniques may require expertise of the data.
The second characteristics of time-series data is that it is not certain
3
that there are enough information available to understand the process. For
example, in electronic nose data, where an array of sensors with various
selectivity for a number of gases are combined to identify a particular smell,
there is no guarantee that the selection of sensors actually are able to identify
the target odour. In nancial data when observing a single stock, which only
measures a small aspect of a complex system, there is most likely not enough
information in order to predict the future (Fama, 1965).
Further, time-series have an explicit dependency on the time variable.
Given an input
x(t)
at time
t
, the model predicts
y(t)
, but an identical input
at a later time could be associated with a dierent prediction. To solve this
problem, the model either has to include more data input from the past or
must have a memory of past inputs. For long-term dependencies the rst ap-
proach could make the input size too large for the model to handle. Another
challenge is that the length of the time-dependencies could be unknown.
Many time-series are also non-stationary, meaning that the characteristics
of the data, such as mean, variance, and frequency, changes over time. For
some time-series data, the change in frequency is so relevant to the task that it
is more benecial to work in the frequency-domain than in the time-domain.
Finally, there is a dierence between time-series data and other types of
data when it comes to invariance. In other domains, for example computer
vision, it is important to have features that are invariant to translations,
rotations, and scale. Most features used for time-series need to be invariant
to translations in time.
In conclusion, time-series data is high-dimensional and complex with
unique properties that make them challenging to analyze and model. There
4
剩余62页未读,继续阅读
资源评论
rui900207
- 粉丝: 0
- 资源: 1
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
最新资源
- 高等数学第一章第二节数列的极限
- Python 版冒泡排序算法源代码
- tensorflow-gpu-2.7.2-cp38-cp38-manylinux2010-x86-64.whl
- tensorflow-2.7.3-cp39-cp39-manylinux2010-x86-64.whl
- tensorflow-2.7.2-cp39-cp39-manylinux2010-x86-64.whl
- Python版本快速排序源代码
- Python 语言版的快速排序算法实现
- 450815388207377安卓_base.apk
- 超微主板 X9DRE-TF+ bios 支持 nvme启动
- 基于Python通过下载气象数据和插值拟合离散数据曲线实现对寒潮过程的能量分析
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
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