Class-Specific Explainability for Deep Time Series Classifiers Ramesh Doddaiah

2025-04-24 0 0 1.45MB 10 页 10玖币
侵权投诉
Class-Specific Explainability for Deep Time Series
Classifiers
Ramesh Doddaiah§
Data Science
WPI
rdoddaiah@wpi.edu
Prathyush Parvatharaju§
Data Science
WPI
psparvatharaju@wpi.edu
Elke Rundensteiner
Data Science, Computer Science
WPI
rundenst@wpi.edu
Thomas Hartvigsen
CSAIL
MIT
tomh@mit.edu
Abstract—Explainability helps users trust deep learning solu-
tions for time series classification. However, existing explainability
methods for multi-class time series classifiers focus on one class
at a time, ignoring relationships between the classes. Instead,
when a classifier is choosing between many classes, an effective
explanation must show what sets the chosen class apart from the
rest. We now formalize this notion, studying the open problem
of class-specific explainability for deep time series classifiers, a
challenging and impactful problem setting. We design a novel
explainability method, DEMUX, which learns saliency maps for
explaining deep multi-class time series classifiers by adaptively
ensuring that its explanation spotlights the regions in an input
time series that a model uses specifically to its predicted class.
DEMUX adopts a gradient-based approach composed of three
interdependent modules that combine to generate consistent,
class-specific saliency maps that remain faithful to the classifier’s
behavior yet are easily understood by end users. Our experimen-
tal study demonstrates that DEMUX outperforms nine state-of-
the-art alternatives on five popular datasets when explaining two
types of deep time series classifiers. Further, through a case study,
we demonstrate that DEMUX’s explanations indeed highlight
what separates the predicted class from the others in the eyes of
the classifier.
I. INTRODUCTION
Background
. Deep learning methods represent the state-
of-the-art for tackling time series classification problems in
important domains from finance [
1
] to healthcare [
2
], [
3
]. In
these applications, we typically aim to classify a time series to
be a member of one of many classes, referred to as multi-class
classification in contrast to binary classification [
4
]. Despite the
success of these deep learning methods, domain experts may
not trust predictions from deep models as they are “opaque”
and thus hard to understand. This lack of trust has the potential
to hinder wide deployment of promising deep models for real-
world applications [
5
]. By helping users better understand and
thus trust their models, explainability has been recognized as
a critical tool for successfully deploying deep learning models
[6], [7], [8], [9], [10], [11], [12], [13].
To explain a time series classification, we highlight the time
steps that the model associates with the class it predicts via a
saliency map [
14
], [
9
], [
7
]. However, in the multi-class problem
setting where a classifier is choosing between many classes,
an effective explanation must show what sets the chosen class
apart from the rest of the classes. Thus the explanation should
§Both authors contributed equally to this research.
ECG Time Series
Multi-class Time
Series Explainer
Normal Heart
Cardiac Arrest
Heart Attack
Normal Heart
Cardiac Arrest
Multi-class Prediction
Class-Specific evidence
relative to other classes
Top Predicted class:
Heart Attack
Top Predicted class:
Heart Attack
Per-class explanations with
overlapping evidences across classes
Multi-class Deep
Time Series Classifier
80%
15%
5%
Fig. 1: Effective explanations of multi-class time series classi-
fiers should show what regions of a time series set that class
apart from the rest. Here, we show a deep classifier that predicts
Heart Attack given an ECG. Explanations for individual classes
(middle column) may have overlapping regions. However, as
shown on right, the middle of the signal is what sets Heart
Attack apart according to the classifier.
only highlight that particular subset of time steps that explains
why that class was predicted compared to other classes. To
further illustrate this with an example from computer vision,
consider classifying images as either cats, dogs, foxes, or
wolves. Fur is certainly evidence that an animal is in the image,
but given choices between only these animals, fur is specific to
no class. Therefore, an explanation should avoid highlighting
fur [
15
]. In computer vision, it has been recognized as critical to
leverage relationships between classes [
15
]. We conjecture that
the same is equally true for time series, yet existing methods
for time series ignore all relationships between classes [
6
], [
9
],
[
8
], [
10
], [
12
]. Further, by ignoring this class-specificity issue,
users may be prone to rationalize a model’s prediction in the
face of erroneous explanations, having been instilled with a
false sense of confidence [16].
Motivating Example.
Consider a deep multi-class time
series classifier for ECG data [
17
] that classifies heartbeats into
three classes: Heart Attack,Cardiac Arrest, and Normal Heart.
Without knowing whether the classifier learned to recognize
relevant critical regions of the time series for each class, a
doctor may not trust its prediction. As shown in Figure 1
(middle), state-of-the-art methods would explain each class
independently [
8
], [
7
], [
13
], [
12
]. However, by ignoring the
1
arXiv:2210.05411v1 [cs.LG] 11 Oct 2022
relationships between classes, it appears as if the model uses
most of the signal when predicting its top class, Heart Attack,
even when it shares the same critical regions with Cardiac
Arrest and Normal Heart classes as shown in blue. A class-
specific explanation, on the other hand, should correct for
the model using the same regions for multiple classes. In our
example, it (rightmost column) explains that the burst (non-
overlapping critical region) of the signal is what the model
uses to predict Heart Attack instead of other classes.
State-of-the-Art.
Explainablility for time series models has
recently emerged as a promising direction to help users trust
deep time series classifier models [
8
], [
9
], [
7
], [
6
], [
10
], [
18
],
[
12
]. The most successful methods learn to perturb input time
series to explain an opaque model’s behavior in the vicinity of
one instance [
7
], [
8
]. Intuitively, time steps that have a higher
impact on model accuracy will be ranked higher. Most existing
methods [
8
], [
9
], [
18
], [
12
], [
13
], [
5
] explain model behavior by
perturbing each time step using either static, predefined values
like zero or other time series instances from a “background”
dataset. For example, PERT [
7
], which explains only binary
deep time series classifiers, perturbs each time step by replacing
it from a replacement time series sampled from the background
dataset. DYNAMASK [
8
], [
19
], which also treats each class
independently, uses static replacement strategies for each time
step for deriving explanations for a multivariate classifier. It
makes a binary decision if a feature is important or not. To-date,
class-specific explanations, despite their recognized need in
fields like computer vision [
15
], remain an open problem in
time series. Typically, a successful multi-class classifier assigns
high probability to one of the classes and lower probabilities
to the rest. Evidence derived to explain the predicted class
should be unique to that class, relative to other classes. But
existing time series explainable methods fail to incorporate the
knowledge about relationships between classes.
Beyond lacking class-specificity, another well-known disad-
vantage of perturbation-learning methods is the high variance
between explanations derived over multiple runs for the same
time series instance as input [
20
], [
19
], [
21
]. High variability
among explanations decreases a user’s trust in an explainability
method and must therefore be reduced.
Problem Definition.
We study the open problem of Class-
Specific Explainability for Multi-Class Time Series Classifiers:
given a time series and a pre-trained multi-class classifier, we
aim to generate a class-specific saliency map for the classifier’s
predicted class. A saliency map is a vector with one element
per time step in the time series instance, where higher values
of an element indicates a higher importance of this time step
according to the classifier. To be class-specific, the saliency
map should assign high importance only to time steps uniquely
important to the predicted class (in contrast to also being
important to other classes). This problem has multiple possibly
conflicting objectives: a good saliency map should be class-
specific, highlight only the most-relevant time steps, and still
remain faithful to the model’s behavior.
Challenges.
Our problem is challenging for several reasons:
•
Class-Specificity: Generating class-specific saliency maps
requires knowledge of explanations across all classes.
However, learning concurrently multiple explanations is
hard, in particular for low-probability classes, with a
model’s predictions often highly variable in regions of
low probability.
•
Local Fidelity: We consider multi-class classifiers that
predict probability distributions. Learning perturbations to
explain these models must incorporate all class probabili-
ties to remain faithful to the classifier’s behavior. However,
minor changes to the input can have a large effect on the
predicted class distribution.
•
Temporal Coherence: Time steps often depend on their
neighbors’ values. This implies that similarly for saliency
maps neighboring time steps should have similar im-
portance. While this encourages discovering important
subsequences, thereby improving explainability, it conflicts
with local fidelity and class-specificity. Hence, a trade-off
must inherently be considered in any effective solution.
•
Consistent Saliency: Perturbation-based explainability
methods can create saliency maps that vary dramatically
for the same instances when re-initialized. Yet to be useful
in real-world applications, we should instead consistently
generate similar explanations for the same time series
Proposed Solution.
To derive class-specific explanations, we
propose
D
istinct T
E
mporal
MU
lticlass E
X
plainer (DEMUX), a
novel model-agnostic, perturbation-based explainability method
for multi-class time series models. DEMUX jointly learns
saliency maps, with a focus on removing shared salient regions
to generate a class-specific explanation for the model’s top
predicted class.
DEMUX is gradient-based approach that monitors changes in
the classifier’s predictions while perturbing values at each time
step. It produces a saliency map for the classifier’s top predicted
class that preserves the classifier’s prediction probability
distribution across classes. To generate good perturbations,
DEMUX learns to sample a replacement time-series per class
from a background dataset using a clustering-based replacement
selector. DEMUX avoids out-of-distribution replacement values
by ensuring perturbations are like other time series the model
has seen before for each class and for each time step, leading
to more stable saliency maps.
Contributions. Our main contributions are as follows:
•
We identify and characterize the problem of class-specific
saliency maps for deep multi-class time series classifiers.
•
We introduce the first effective solution, DEMUX, which
extends beyond recent work with three innovations:
learning to remove shared saliency across classes (Class-
Specificity), generating class-specific perturbations that
are locally faithful (Local Fidelity and Temporal Coher-
ence), and ensuring stability of saliency maps (Consistent
Saliency) for given time series instances.
•
Using five real datasets, we conclusively demonstrate that
DEMUX outperforms nine state-of-the-art alternatives,
successfully generating class-specific explanations for
multiple types of deep time-series classifiers.
2
摘要：

Class-SpecicExplainabilityforDeepTimeSeriesClassiersRameshDoddaiah§DataScienceWPIrdoddaiah@wpi.eduPrathyushParvatharaju§DataScienceWPIpsparvatharaju@wpi.eduElkeRundensteinerDataScience,ComputerScienceWPIrundenst@wpi.eduThomasHartvigsenCSAILMITtomh@mit.eduAbstract—Explainabilityhelpsuserstrustdeepl...

展开>> 收起<<
Class-Specific Explainability for Deep Time Series Classifiers Ramesh Doddaiah.pdf

共10页,预览2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！

相关推荐

更多
立即下载
分类：图书资源 价格：10玖币 属性：10 页 大小：1.45MB 格式：PDF 时间：2025-04-24

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务

作者详情

相关内容

更多

热门标签

人际关系 动力学连接体 力的合成 配电装置 高考理综 全宋诗作者索引 公务员考试
/ 10
评分 收藏
分享
立即下载
客服
关注