Empirical Evaluation of Data Augmentations for Biobehavioral Time Series Data with Deep Learning Huiyuan Yang Han Yu and Akane Sano

2025-04-24 0 0 1.55MB 12 页 10玖币

侵权投诉

Empirical Evaluation of Data Augmentations for Biobehavioral Time Series Data

with Deep Learning

Huiyuan Yang, Han Yu and Akane Sano

Department of Electrical Computer Engineering

Rice University, Houston TX 77005, USA

{hy48, hy29, Akane.Sano}@rice.edu

Abstract

Deep learning has performed remarkably well on many

tasks recently. However, the superior performance of deep

models relies heavily on the availability of a large number

of training data, which limits the wide adaptation of deep

models on various clinical and affective computing tasks,

as the labeled data are usually very limited. As an effec-

tive technique to increase the data variability and thus train

deep models with better generalization, data augmentation

(DA) is a critical step for the success of deep learning mod-

els on biobehavioral time series data. However, the effec-

tiveness of various DAs for different datasets with differ-

ent tasks and deep models is understudied for biobehav-

ioral time series data. In this paper, we ﬁrst systematically

review eight basic DA methods for biobehavioral time se-

ries data, and evaluate the effects on seven datasets with

three backbones. Next, we explore adapting more recent DA

techniques (i.e., automatic augmentation, random augmen-

tation) to biobehavioral time series data by designing a new

policy architecture applicable to time series data. Last, we

try to answer the question of why a DA is effective (or not)

by ﬁrst summarizing two desired attributes for augmenta-

tions (challenging and faithful), and then utilizing two met-

rics to quantitatively measure the corresponding attributes,

which can guide us in the search for more effective DA for

biobehavioral time series data by designing more challeng-

ing but still faithful transformations. Our code and results

are available at Link.

1. Introduction

Deep learning performs remarkably well in many ﬁelds,

including computer vision (CV), natural language process-

ing (NLP), and recently time series-related tasks [6,10,

32]. Those successful applications increasingly inspire re-

searchers to embrace deep learning for solving issues in

human-centered applications that use physiological and be-

havioral time series data. However, the superior perfor-

mance of deep models relies heavily on the availability of a

large number of training data, but unfortunately, many hu-

man centered applications (i.e., healthcare tasks) usually do

not have enough labeled samples, which may limit the wide

adaptation of deep models to various computing tasks.

As an effective technique to increase the data variability

and thus train deep models with better generalization, data

augmentation (DA) is a critical step for the successful appli-

cations of deep learning models. While DA can yield con-

siderable performance improvements, they do require do-

main knowledge and are task- and domain-dependent. For

example, image rotation, a likely class-preserving behavior,

is designed to rotate the input by some number of degrees.

The image’s class can still be recognized by humans, thus

allowing the model to generalize in a way humans expect

it to generalize. However, such an effective random angle-

based rotation operation may not be applicable to other do-

mains, i.e., wearable data. In addition, searching for the

most effective DA methods for a new dataset is very time-

consuming, and this motivated the proposal of several auto-

matic DA search algorithms [4,5,14–16].

The existing DA literature mainly focuses on computer

vision, but its application to other domains, i.e, biobehav-

ioral time series data, is understudied. A few works inves-

tigated the effectiveness of basic DA methods for time se-

ries and wearable data [1,11,30,32]. However, those works

only investigated the very basic DAs, leaving the more re-

cent DA techniques (i.e., automatic DA) unexplored. More

importantly, it is still an open question of why a DA method

works, and how to quantify its effectiveness. Therefore,

in this paper, we ﬁrst systematically review various basic

DA methods for biobehavioral time series data, evaluating

the effects on different datasets with varied backbones and

tasks. Next, we validate the effectiveness of adapting more

recent DA techniques (i.e., automatic DA) to biobehavioral

time series data. Following the DADA [14], we designed a

different policy architecture where the operations are differ-

entiable with respect to different time series DA methods.

arXiv:2210.06701v1 [cs.LG] 13 Oct 2022

Figure 1. Examples of different data augmentation methods used in the experiments. The red lines indicate the input data and the green lines

are the augmented data based on the eight data augmentation operations including Jittering, Scaling, Rotation, Permutation, Magnitude

Warping, Time Warping, and Window Warping with two different magnitudes.

Therefore, the model can be applied to biobehavioral time

series data, and the DA parameters and deep model weights

can be jointly optimized. Lastly, we try to answer the open

question of why a DA works(or not), by ﬁrst summarizing

two desired attributes (challenging and faithful) for an ef-

fective DA, and then utilizing two metrics to quantitatively

measure the two attributes. We ﬁnd that an effective DA

needs to generate challenging but still faithful transforma-

tions, which can guide us for the search of more effective

DA for biobehavioral time series data. The contributions of

this work are summarized as follows:

• A comprehensive and systematic evaluation of eight

data augmentation methods on seven biomedical time

series datasets with three backbones for different

tasks.

• We revisit the automatic DA methods to make the op-

erations are differentiable with respect to different time

series DA methods, therefore can be applied to biobe-

havioral time series data. Besides, random DA is also

investigated to boost efﬁciency.

• We summarize two desired attributes for an effective

DA, and adopt two metrics to quantitatively measure

the two attributes respectively. Recommendations are

summarized for the search of more effective data aug-

mentation methods.

2. Related Works

2.1. Augmentations for Biobehavioral Time Series

Data

Most of the basic DA methods are borrowed or inspired

from image or time series data augmentation, such as ﬂip-

ping, cropping and noise addition. These augmentation

methods rely on adding random transformations to the train-

ing data. Um et al. [30] systematically evaluated six DA

methods for wearable sensor data based Parkinson’s disease

monitoring, and found that the combination of rotational

and permutational data augmentation methods improve the

baseline performance the most. Ohashi et al. [20] proposed

a rotation based data augmentation method for wearable

data, which can take the physical constraint into account.

Alawneh et al. [1] investigated the beneﬁts of adopting time

series data augmentation methods to biomedical time series

data, and demonstrated the improved accuracy of several

deep learning models for human activity recognition. Ey-

obu and Han [27] proposed an ensemble of feature space

augmentation methods, which was used for human activity

classiﬁcation based on wearable inertial measurement unit

(IMU) sensors. Besides, DA methods have been also used

to balance the dataset. For example, Cao et al. [3] used DA

methods to balance the number of samples among different

categories for automated heart disease detection. However,

those related works only investigated the very basic DAs,

and the effectiveness of adapting more advanced DAs is not

explored yet for biobehavioral time series data. More im-

portantly, the previous works did not explore the question

why a DA is effective, and vice versa.

2.2. Automatic Data Augmentation

DA can be very useful for the training of deep models,

but the success relies heavily on domain knowledge and also

the extensive experiments to select the effective DA poli-

cies for a target dataset. Otherwise, a model may be neg-

atively impacted by some DA policies [11,32]. Therefore,

it is nontrivial and desired to select the effective DA policy

for a new dataset automatically. The goal of automatic data

augmentation is to search for effective data augmentation

policies that, when applied during the model training, will

minimize its validation loss, therefore better generalization

ability. The pioneering work, AutoAugment [4], formats

the process of searching DA as an optimization problem

to search for the parameters of augmentation, and follwing

work [5,9,14–16] were later proposed to improve the efﬁ-

ciency. Despite the success of automatic DA for computer

vision tasks, the adaption to biobehavioral time series data

is understudied. The only work we know is [24], which

investigated the automatic differentiable data augmentation

for EEG signals. However, our work is different with [24],

as we target more diverse types of data and DA methods,

and more importantly, we explore to explain why a DA

works or not.

Figure 2. Examples of time series data augmented by two consec-

utive operations. The ﬁrst column is the original input signal, with

two random operations sequentially applied to the input signal, the

ﬁnal augmented signal is shown in the third column (as green line).

Note that, not only the name of operations and magnitude matter,

but also the order of operation.

2.3. Quantitative Measurement of Effectiveness for

Data Augmentation

Although the effectiveness of DA is well acknowledged,

a quantitative evaluation of the effectiveness of DA is still an

open question. Currently, the most well-known hypothesis

is that effective DA can produce samples from an ”overlap-

ping but different” distribution [2,17], therefore improving

generalization by training with the diverse samples. How-

ever, the role of distribution shift in training remains un-

clear. A more recent work [8] studied to quantify how DA

improves model generalization, and introduced two mea-

sures: Afﬁnity and Diversity, to predict the performance

of an augmentation method. During our experiments, we

adopt those two metrics to jointly evaluate the two attributes

of different augmentations.

3. Methods

We performed extensive experiments with various DA

methods on different datasets with different tasks and back-

bones. After that, an automatic DA search strategy is

adapted to jointly optimize the deep models and also the

best DA policies for wearable data. To avoid the compli-

cated searching procedure of automatic DA while keeping

the effectiveness, a random augmentation procedure was

then adapted in our experiments. Lastly, We also explored

to answer the open questions of why some DA methods are

more effective than others by quantitatively measuring the

effectiveness.

3.1. Basic Data Augmentation Methods

DA methods rely on adding random transformations to

the training data, and the transformations can be gener-

ally classiﬁed into three categories: magnitude-based, time-

based, and frequency-based transformations. Magnitude-

based transformations are applied to the wearable data

along the variate or value axes. Time-based transforma-

tions change the time steps, and frequency-based transfor-

mations warp the frequencies, respectively. During our ex-

periments, we will mainly focus on the magnitude and time-

based transformations.

Let Tαdenote an augmentation operations parameterized

by α, given input data x, this procedure outputs augmented

data ˆx=Tα(x), where αcontrols the magnitude of the

operation T. We consider a set of DA methods drawn from

the traditional time series processing literature. Speciﬁcally,

our augmentation set consists of eight operations: Jittering,

Scaling, Rotation, Permutation, Magnitude Warping, Time

Warping, Window Slicing, and Window Warping.

Given an input data x= [x1, x2, . . . , xT]with T, the

number of time steps, and each element xtcan be univariate

or multivariate.

Jittering. A random noise is added to the input data:

ˆx =Tα(x)=[x1+1, x2+2, . . . , xT+T](1)

where ˆx is the augmented data, is a random noise (i.e.,

Gaussian noise) added to each time step t. Assume ∼

N(0, σ2), and the standard deviation α={σ}of the added

noise is a hyper-parameter that needs to be pre-determined.

Scaling. The magnitude of the data in a window is changed

by multiplying a random scalar.

ˆx =Tα(x) = [x1, x2, . . . , xT](2)

where can be determined by a Gaussian distribution ∼

N(1, σ2)with α={σ}as a hyperparameter.

Rotation. Rotate each element by a random rotation matrix.

Although rotating data by a random angle can create plau-

sible patterns for images, it might not be suitable for time

series data. A widely used alternative is ﬂipping, which is

deﬁned as:

ˆx =Tα(x)=[−x1,−x2,...,−xT](3)

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

EmpiricalEvaluationofDataAugmentationsforBiobehavioralTimeSeriesDatawithDeepLearningHuiyuanYang,HanYuandAkaneSanoDepartmentofElectricalComputerEngineeringRiceUniversity,HoustonTX77005,USAfhy48,hy29,Akane.Sanog@rice.eduAbstractDeeplearninghasperformedremarkablywellonmanytasksrecently.However,thesuper...

展开>> 收起<<

Empirical Evaluation of Data Augmentations for Biobehavioral Time Series Data with Deep Learning Huiyuan Yang Han Yu and Akane Sano.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Empirical Evaluation of Data Augmentations for Biobehavioral Time Series Data with Deep Learning Huiyuan Yang Han Yu and Akane Sano

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: