Scribd
Upload
26 views

4 An Experimental Comparison of Two Machine Learning Approaches

1) The document compares two machine learning approaches for classifying emotions from facial expressions: one approach trains a single model on data from both males and females, while the other trains separate models for each gender. 2) Preliminary results show that the approach with separate gender models produces more accurate classifications than the single combined model. However, the separate male model currently achieves higher accuracy than the female model. 3) The researchers aim to improve the analysis by increasing their sample sizes and performing significance testing on the results. They hypothesize that accounting for gender differences can enhance the accuracy of emotion classification using machine learning.

Uploaded by

Hub Technology
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
26 views

4 An Experimental Comparison of Two Machine Learning Approaches

1) The document compares two machine learning approaches for classifying emotions from facial expressions: one approach trains a single model on data from both males and females, while the other trains separate models for each gender. 2) Preliminary results show that the approach with separate gender models produces more accurate classifications than the single combined model. However, the separate male model currently achieves higher accuracy than the female model. 3) The researchers aim to improve the analysis by increasing their sample sizes and performing significance testing on the results. They hypothesize that accounting for gender differences can enhance the accuracy of emotion classification using machine learning.

Uploaded by

Hub Technology
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

Emotion Classification using Machine learning

An Experimental Comparison of Two


Machine Learning Approaches for Emotion
Classification
Emergent Research Forum Paper
Wangchuchu Zhao Keng Siau
Missouri University of Science and Missouri University of Science and
Technology Technology
[email protected] [email protected]
Abstract
Correctly identifying an emotion has always been challenging for humans, not to mention machines! In
this research, we use machine learning to classify human emotion. Emotional differences between
genders are well documented in fields like psychology. We hypothesize that genders will impact the
accuracy of classifying emotion with machine learning. Two different machine learning approaches were
tested in an experimental study. In one approach, emotions from both genders were used to train the
machine. In another approach, the genders were separated and two separate machines were used to learn
the emotions of the two genders. Our preliminary results show that the approach where the genders were
separated produces higher accuracy in classifying emotion.

Keywords

Emotion classification, facial expression, sexes, machine learning.

Introduction
Mining information from data to support decision making is increasing viewed as a competitive necessity
(Lee and Siau 2001). Emotion classification is a form of information mining (Zhao and Siau 2017) and
emotion classification has been increasingly used in varied domains, especially in multimedia, computer
science, data science, and marketing (Adeborna and Siau 2014). In computer science, detecting and
modeling emotion plays an important role in Human-Computer interactions, and feedback from the
users’ emotions can be used to improve user experience (Peng et al. 2014). In the multimedia domain,
emotion classification on affective characters in response to video are studied (Soleymani et al. 2012). In
the marketing field, emotion classification is important in measuring public opinions and analyzing
customer performances (Rao 2016). Correctly identifying an emotion is important for relationship
building and networking in both business and social settings. Because of the importance of emotion
classification, machine learning approaches to emotion classification have attracted the attention of many
researchers. The potential payoff can be huge if machines can classify different emotions more accurately
and provide real-time information for decision makers. Compared to traditional sentiment analysis based
on comments, emotion classification based on facial expression can accurately reflect perceptual
sentiment of customers. Real-time emotion classification can also project a graph of customer sentiment
over a period of time.
For example, real-time emotion analysis of facial expressions has been utilized during the Clinton vs
Trump presidential debates using the emotion API software developed by Microsoft. As can be seen in
Figure 1, the emotions detected by the software are extremely different for the two candidates. Though
the real reason may vary and may be unique to each individual to some extent, it could also be attributed
to sex difference. The software does not distinguish between the male and female groups during the
analysis, which may affect the precision of the analysis. In this research, we hypothesize that separating
the two genders during the machine learning phase in emotion classification process will produce a more
accurate analysis of emotion.

Twenty-third Americas Conference on Information Systems, Boston, 2017 1


Emotion Classification using Machine learning

Figure 1. Real-time Emotion Analytics of Clinton vs Trump 3rd Debate


(http://www.economist.com/blogs/graphicdetail/2016/10/daily-chart-12)

Research Methodology and Procedure


The research model is shown in Figure 2. Approach A represents the existing techniques to classifying
emotions. Approach B is the approach proposed in this research. In Approach A, the machine learning
algorithm is trained using both male and female facial expressions and emotions. In Approach B, we
suggest that the subjects’ genders should be identified first and then the appropriate gender’s emotion
algorithm is applied based on the gender.
For this research, a standard database is used to train the machines on the standard emotions. We used
80 percent of the data as training data, and 20 percent of the data to test the accuracies of emotion
classification.

Emotion H1
Approach A recognition Emotion
algorithm (male
and female)

H2
Male emotion
Gender expression Emotion
Approach B identifica algorithm
tion H3
algorithm Female emotion
expression
Emotion
algorithm

Figure 2. Research Model

Twenty-third Americas Conference on Information Systems, Boston, 2017 2


Emotion Classification using Machine learning

Preliminary Results
The preliminary results are summarized in Table 1. Table 1 shows that Approach B (i.e., 79% and 82%) is
better than Approach A (i.e., 67%). When genders are separated, the machine learning approach provides
better emotion recognition accuracies than when the genders are combined (i.e., facial expressions of both
male and female subjects). However, the results seem to show that the machines are identifying emotion
of male subjects better than female subjects, which is contrary to our expectation.
We are currently performing more analyses using the data:
(i) We are increasing the sample size for each set (i.e., combined faces, female faces, and male faces)
(ii) We are doing significance testing

Result of the
whole database
Combined frontal 67%
faces
Female frontal 79%
faces
Male frontal faces 82%

Table 1. Evaluation Results

Conclusions and Expected Contributions


We hypothesize that gender will impact the accuracies of emotion classifications and two approaches were
experimentally tested. Our preliminary results show that there are indeed differences between the two
genders. Our preliminary results also show that emotion classification is more accurate when the male
and female sets are separated than when the sets are combined (i.e., both male and female faces in one
set).
Machine learning approaches to classifying emotions are important in many areas of business (e.g., retail
sales, restaurants) and in other domains as well (e.g., airport security). Our study contributes not only to
research in enhancing the accuracy of classifying emotions based on machine learning approaches, but
also to practitioners that are trying to use big data and machine learning approaches to improve decision-
making and business performance.

REFERENCES
Adeborna, E., and Siau, K. 2014. "An Approach to Sentiment Analysis-the Case of Airline Quality Rating,"
The Pacific Asia Conference on Information Systems (PACIS), p. 363.
Lee, S., and Siau, K. 2001. "A Review of Data Mining Techniques," Industrial Management & Data
Systems (101:1), pp. 41-46.
Peng, Y., Zhu, J.-Y., Zheng, W.-L., and Lu, B.-L. 2014. "Eeg-Based Emotion Recognition with Manifold
Regularized Extreme Learning Machine," Engineering in Medicine and Biology Society (EMBC),
2014 36th Annual International Conference of the IEEE: IEEE, pp. 974-977.
Rao, Y. 2016. "Contextual Sentiment Topic Model for Adaptive Social Emotion Classification," IEEE
Intelligent Systems (31:1), pp. 41-47.
Soleymani, M., Pantic, M., and Pun, T. 2012. "Multimodal Emotion Recognition in Response to Videos,"
IEEE transactions on affective computing (3:2), pp. 211-223.

Twenty-third Americas Conference on Information Systems, Boston, 2017 3


Emotion Classification using Machine learning

Zhao, W., Siau, K. 2017. “Machine Learning Approaches to Sentiment Analytics,” Twelve Annual Midwest
Association for Information Systems Conference (MWAIS 2017), Springfield, Illinois.

Twenty-third Americas Conference on Information Systems, Boston, 2017 4

You might also like