To Improve Is to Change Towards Improving Mood Prediction by Learning Changes in Emotion

2025-05-06 0 0 3.79MB 6 页 10玖币

侵权投诉

To Improve Is to Change: Towards Improving Mood Prediction

by Learning Changes in Emotion

Soujanya Narayana

University of Canberra

ACT, Australia

soujanya.narayana@canberra.edu.au

Ramanathan Subramanian

University of Canberra

ACT, Australia

ram.subramanian@canberra.edu.au

Ibrahim Radwan

University of Canberra

ACT, Australia

ibrahim.radwan@canberra.edu.au

Roland Goecke

University of Canberra

ACT, Australia

roland.goecke@ieee.org

Valence

Emotion change

Mood label:

"Negative"

Time

-1

Figure 1: Problem Illustration: Figure depicts emotion changes in an input video sample having a negative mood label. The

top colour bar denotes per-frame valence values for the video, while the bottom colour bar depicts emotional valence change

(Δ) labels over a window of ve frames (best viewed in colour).

ABSTRACT

Although the terms mood and emotion are closely related and often

used interchangeably, they are distinguished based on their dura-

tion, intensity and attribution. To date, hardly any computational

models have (a) examined mood recognition, and (b) modelled the

interplay between mood and emotional state in their analysis. In

this paper, as a rst step towards mood prediction, we propose a

framework that utilises both dominant emotion (or mood) labels,

and emotional change labels on the AFEW-VA database. Experi-

ments evaluating unimodal (trained only using mood labels) and

multimodal (trained with both mood and emotion change labels)

convolutional neural networks conrm that incorporating emo-

tional change information in the network training process can

signicantly improve the mood prediction performance, thus high-

lighting the importance of modelling emotion and mood simulta-

neously for improved performance in aective state recognition.

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 prot 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 the

author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or

republish, to post on servers or to redistribute to lists, requires prior specic permission

and/or a fee. Request permissions from permissions@acm.org.

ICMI ’22 Companion, November 7–11, 2022, Bengaluru, India

ACM ISBN 978-1-4503-9389-8/22/11. . . $15.00

https://doi.org/10.1145/3536220.3563685

CCS CONCEPTS

•Human-centered computing →Ambient intelligence

;

•

Computing methodologies →Supervised learning by clas-

sication;•Applied computing →Psychology.

KEYWORDS

Mood; Emotion; Convolution neural network; Unimodal; Multi-

modal

ACM Reference Format:

Soujanya Narayana, Ramanathan Subramanian, Ibrahim Radwan,

and Roland Goecke. 2022. To Improve Is to Change: Towards Improving

Mood Prediction by Learning Changes in Emotion. In INTERNATIONAL

CONFERENCE ON MULTIMODAL INTERACTION (ICMI ’22 Companion),

November 7–11, 2022, Bengaluru, India. ACM, New York, NY, USA, 6 pages.

https://doi.org/10.1145/3536220.3563685

1 INTRODUCTION

There is mounting evidence that emotions play an essential role in

rational and intelligent behaviour. Besides contributing to a richer

quality of interaction, they directly impact a person’s ability to

interact in an intelligent way [

]. Quite often, the terms mood and

emotion are used interchangeably, although dierences in duration,

intensity and attribution exist. While

emotion

is a short-term af-

fective state induced by a source which can last for a few minutes,

mood

refers to a longer-term aective state that can last for hours

or even days, and be without a causal source [

]. Most research

in aective computing has focused on inferring emotional states,

arXiv:2210.00719v1 [cs.HC] 3 Oct 2022

ICMI ’22 Companion, November 7–11, 2022, Bengaluru, India S. Narayana, R. Subramanian, I. Radwan, R. Goecke

Mood

classification

Single branch CNN

(mood labels only)

Model-fusion

(mood and delta

labels)

Two-branch CNN

(mood and delta

labels)

Teacher-Student

Model

Input

Figure 2: Overview of the proposed mood recognition framework.

while very little research has so far been devoted to automated

mood recognition [

] or the joint modelling of the interplay be-

tween emotion and mood for improved aective state recognition.

Psychological studies on mood have made substantial progress.

An eye-tracking study has revealed that positive mood results in

better global information processing than a negative mood [

The authors in [

] have observed a mood-congruity eect, where

positive mood hampers the recognition of mood-incongruent nega-

tive emotions and vice-versa. The mood-emotion loop is a theory

that posits mood and emotion as distinct mechanisms, which aect

each other repeatedly and continuously. This theory argues that

mood is a high-level construct activating latent low-level states

such as emotions [

]. Recognising the interactions between mood

and emotion has the potential to lead to a better understanding

of aective phenomena, such as mood disorders and emotional

regulation.

On the contrary, mood recognition has rarely been addressed

from a computational perspective and only a few studies have

explored mood [

]. Body posture and movement correlates of

mood have been explored in [

]. User mood is induced via musical

stimuli and the authors have observed that head posture and move-

ments characterise happy and sad mood. Katsimerou et al. [

] have

examined automatic mood prediction from recognised emotions,

showing that clustered emotions in the valence-arousal space pre-

dict single moods much better than multiple moods within a video.

Research on mood prediction has also neglected to investigate the

interplay between mood and emotion, though the psychological

literature recognises a relationship between the two [20].

From an aective computing viewpoint, developing a mood

recognition framework requires ground-truth mood labels for

model training, but only very few databases record the user mood

(directly or indirectly via an observer). Widely used aective cor-

pora, such as AFEW-VA [

], HUMAINE [

], SEMAINE [

] and

DECAF [

] only contain dimensional and/or categorical emotion

labels. One of the few datasets with mood ratings is EMMA [

where the annotations developed represent the overall emotional

impression of the human annotator (or observer) for the examined

stimulus [

]. Machine learning approaches have been extensively

used for inferring emotions from visual, acoustic, textual and neu-

rophysiological data [

]. Contemporary studies em-

phasise the improved performance of multimodal approaches to

the detection of emotional states vis-á-vis unimodal ones [

]. Re-

cent studies characterise mood disorders, such as depression, by

examining speech style, eye activity, and head pose [

]. Deng

et al. [

] propose a multitask emotion recognition framework that

can deal with missing labels employing a teacher-student paradigm.

Knowledge Distillation (KD) is a technique that enables the trans-

fer of knowledge between two neural networks, unifying model

compression and learning with privileged information [

]. KD

techniques have been employed for facial expression recognition

where the teacher has access to a fully visible face, whereas the

student model only has access to occluded faces [10].

While our research is ultimately aimed towards mood prediction

and understanding the interplay between mood and emotions from

video data, the present study is an initial step on this path. We

use the AFEW-VA dataset to derive (a) dominant emotion labels,

which refer to the emotion persisting for most consecutive frames

(termed mood labels), and (b)

or emotion change labels, which

represent the change in emotion over a xed window size. Given

the sparsity of in-the-wild data with mood annotations and the pre-

liminary nature of this study, the dominant emotion labels are used

here in lieu of actual mood labels. In the future, we will be using

actual mood labels derived from expert annotators. Fig. 1 illustrates

how emotion change is captured for an exemplar video clip, while

Fig. 2 overviews our dominant emotion or mood prediction frame-

work. A unimodal 3D Convolutional Network Network (3D CNN)

is trained using only mood labels, while a two-branch (multimodal)

CNN model, multi-layer perceptron, and a teacher-student model

are evaluated for fusing emotion-mood information for mood pre-

diction. Empirical evaluation reveals that incorporating emotion

change information improves mood prediction performance by as

much as 54%, conrming the salience of ne-grained emotional

information for coarse-grained mood prediction. This study makes

the following contributions:

•

To the best of our knowledge, from a computational mod-

elling perspective, this is the rst study to examine mood

prediction incorporating both mood and emotional infor-

mation. Mood labels are derived from valence annotations,

instead of subjective impressions provided by a human an-

notator.

•

The experimental evaluation of multiple models shows that

incorporating emotional change information is benecial

and can produce a signicant improvement in mood predic-

tion performance.

2 MATERIALS

2.1 Dataset

Here, the AFEW-VA [

] dataset, a subset of the AFEW [

], com-

prising 600 video clips extracted from feature lms at a rate of 25

frames per second, was used. Video clips in this dataset range from

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：

ToImproveIstoChange:TowardsImprovingMoodPredictionbyLearningChangesinEmotionSoujanyaNarayanaUniversityofCanberraACT,Australiasoujanya.narayana@canberra.edu.auRamanathanSubramanianUniversityofCanberraACT,Australiaram.subramanian@canberra.edu.auIbrahimRadwanUniversityofCanberraACT,Australiaibrahim.rad...

展开>> 收起<<

To Improve Is to Change Towards Improving Mood Prediction by Learning Changes in Emotion.pdf

共6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！

To Improve Is to Change Towards Improving Mood Prediction by Learning Changes in Emotion

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: