affective mood mining through deep recurrent neural...
TRANSCRIPT
Affective Mood Mining Through Deep Recurrent Neural Network
Md. Golam Rabiul Alam, Sarder Fakhrul Abedin, Seung Il Moon, Ashis Talukder, Anupam Kumar
Bairagi, Md. Shirajum Munir, Saeed Ullah , Choong Seon Hong
Department of Computer Science and Engineering, Kyung Hee University
{robi, saab0015, moons85,ashis, anupam, munir, saeed}@khu.ac.kr, [email protected]
Abstract
Affective computing is becoming a pioneer research domain to understand human’s emotion through scientific
methods. From genome sequence to face recognition and from neuroimaging to social-post mining each of this
domain tries to use their scientific methodology to recognize, realize and predict human’s affective state. The
pen and paper-based affective state determination methods are not so accurate and impressive therefore due to
the advancement of intelligent technology researchers are trying to apply some intelligent learning methods to
realize individuals affective state. This research uses biosensors data to realize humans’ affective state.
Humans’ psychophysiological data is collected through Electroencephalogram (ECG), Electro-Dermal
Activity (EDA), Electromyography (EMG) and Photoplethysmogram (PPG) and analyzed those data using the
deep recurrent neural network to determine affective mood. Here, based on Russell’s circumplex four primary
affective mood i.e. Joy, Sad, Surprise, and Disgust is considered for realization. The benchmark DEAP dataset
is used to analyze the performance of the proposed method. The higher accuracy in classification of the primary
affective mood justifies the performance of the proposed method.
1. Introduction
The human mind is so complex and still remains unexplored
black-box to the scientific community. According to
neuroimaging research, the brain of a human consists 100 billion
neurons and glial cells and they form a well-structured
communication network among themselves, and therefore
pinpointing the neurons, which are responsible for which mental
state still remains haze [1]. However, one of the pioneering
research by J. D. Haynes et al. [2], disclosed the fact and
feasibility of decoding the mental state of humans such as
conscious experience and covert attitude by evaluating non-
invasive fMRI signals. Apart from that, H. L. Niculescu et al. [3]
successfully classified the blood biomarkers, which is a state-of-
the-art method to recognize suicidality in the patients with mood
disorder. There is some research on affective computing for
determining affective state as an emotional state from face
recognition [4]. The authors of [5] derive the emotional state from
social network post. However, face recognition based, social
network post based or tradition questionnaire based emotion
recognition has many flaws. Because sometimes facial
expression cannot express the internal affective state, individuals’
are not posting on the social network in every time, and people
may hide information in case of questionnaire-based emotion
recognition. Therefore, biosensor’s physiological information is
relatively more appropriate to analyze humans’ affective state.
The DEAP [6] is a pioneer research in analyzing affective state
from physiological data. In that research, the authors’ used 45
channels biosensors data including EEG, EMG, ECG, PPG, GSR
[7] etc. and used simple Naïve Bayes classification to analyze
affective mood. In contrast, we used only four channels of
biosensors i.e. Electroencephalogram (ECG), Electro-Dermal
Activity (EDA), Electromyography (EMG) and
Photoplethysmogram (PPG) to classify affective state. In this
research, we determine the Valence and Arousal level of a human
subject while watching the video stimuli of different affective
state. Based on the Valence and Arousal level, the affective state
is determined according to Russell’s circumplex [] as shown in
Fig. 1.
2. Deep Recurrent Neural Network (RNN) based Affective
Mood Mining
Biosensors signals, which are collected from the subject are
Fig. 1. Simplified Russell’s mood circumplex [3]
2017년 한국소프트웨어종합학술대회 논문집
501
continuous and sequential data signals. There are several machine
learning methods for processing sequential data like HMM,
MEMM, MRF, and CRF. However, the deep recurrent neural
network has the capability to work with the massive dataset.
Therefore, the deep recurrent neural network (RNN) is used to
classify the affective moods of biosensor’s signal.
The human mood mostly depends on the current environment,
facts, stimulus and past emotional states. Therefore, for
predicting affective mood we need to consider the current input
and past states. In RNN model, the long short-time memory
(LSTM) has the capability to memorize the near past and derive
the dependency from the previous states and current biosensors
observations. The used RNN model with LSTM architecture is
presented in Fig. 2.
Fig. 2. LSTM based RNN model for mood mining. The inputs are the biosensor observations and outputs are the
valence and arousal level induced by the video stimuli to determine the mood of the user.
The Long Short-Term Memory (LSTM) helps the model to hold
long-term dependencies. The forget gate layer 𝐹𝑡 determines
what information we’re going to throw away from the cell state
as determine using (1)
The input gate layer it helps to decide which values we’ll update
is derived from (2).
Also, the tanh layer creates a vector of new candidate values
Ĉ𝑡to store in the cell state as derive in (3).
Then update the old cell state 𝐶𝑡−1 to new cell state 𝐶𝑡
according to (4).
Later, we run a sigmoid layer which decides what parts of the cell
state we’re going to output as derived from (5).
Then, we put the cell state through tanh to push the values to be
between -1 to +1, as shown in (6).
Finally, the softmax function as in (7) or normalized exponential
function highlights the largest values and suppress values which
are significantly below the maximum value to determine the
accurate valence and arousal level with highest values.
3. Experimental Results
The deep RNN based affective mood classification method, we
applied in DEAP [6] dataset.
3.1. Dataset Description
The dataset DEAP is an open access (required registration and
authorization) standard dataset used in many emotion analysis
research. To prepare the dataset authors’ were placed 45 recorded
channels at 512Hz, among them 32 EEG channels, 12 peripheral
channels, and 1 status channel. Out of the 45 channel, we used
the data of four channels i.e. Electroencephalogram (ECG),
Electro-Dermal Activity (EDA), Electromyography (EMG) and
𝐶𝑡 = 𝐹𝑡 ∗ 𝐶𝑡−1 + 𝑖𝑡 ∗ Ĉ𝑡 (4)
𝐹𝑡 = 𝜎(𝑊𝑓 . [𝐻𝑡−1, 𝑋𝑡) + 𝑏𝑓 (1)
𝑖𝑡 = 𝜎(𝑊𝑖. [𝐻𝑡−1, 𝑋𝑡) + 𝑏𝑖 (2)
Ĉ𝑡 = 𝑡𝑎𝑛ℎ(𝑊𝐶 . [𝐻𝑡−1, 𝑋𝑡) + 𝑏𝐶 (3)
𝑂𝑡 = 𝜎(𝑊𝑜. [𝐻𝑡−1, 𝑋𝑡) + 𝑏𝑜 (5)
𝐻𝑡 = 𝑂𝑡 ∗ 𝑡𝑎𝑛ℎ(𝐶𝑡) (6)
𝑌𝑇 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑊𝑓 . 𝐻𝑇) (7)
2017년 한국소프트웨어종합학술대회 논문집
502
Photoplethysmogram (PPG). Signals are taken from the subjects’
body surface of the above mentioned 45 channels while watching
the emotional video clips. There were 32 volunteers watched 40
of the music video clips. Using the Likert scale the participants
rated all the video clips during the experiment. Valence, arousal,
dominance, and liking were rated directly after each trial on a
continuous 9-point scale.
The length of each video clip is 2 minutes. Total 8064 samples
from each channel are taken from the attached body sensors while
watching each of the 2 minutes video clip. Therefore, the total
data size for our experiment is 32x40x4x8064 of 32 participants,
40 video clips, 4 channels, and 8064 samples data points.
3.2. Performance study
The classification accuracy of four primary affective moods is
determined in the form of confusion matrix as presented in Table
1. According to the confusion matrix, the overall accuracy is 89%.
Table 1: Confusion Matrix of affective Mood
Classification (Unit: %)
Joy Surprise Sad Distressed
Joy 22 3 0 0
Surprise 2 20 0 0
Sad 0 0 24 2
Distressed 1 2 1 23
Average
Accuracy 88% 80% 96%
92%
Overall
Accuracy 89%
4. Conclusion
Wearable internet of things market is rapidly growing especially
for health status monitoring and athletics training. Affective
mood recognition from wearable biosensors can complement
context aware recommendation, mood stabilization, stress and
depression management, especially for mental well-being. The
accuracy of the proposed deep RNN based affective mood mining
method is better but not impressive. The discriminative feature
extraction and more deep architecture may enhance the accuracy
of the proposed method.
Acknowledgement
This research was supported by the MSIT(Ministry of Science
and ICT), Korea, under the Grand Information Technology
Research Center support program (IITP-2017-2015-0-00742)
supervised by the IITP(Institute for Information &
communications Technology Promotion)" *Dr. CS Hong is the
corresponding author.
References
[1] Alam, Md Golam Rabiul and Cho, Eung Jun and Huh, Eui-
Nam and Hong, Choong Seon, “Cloud-based mental state
monitoring system for suicide risk reconnaissance using wearable
bio-sensors,” Proceedings of the 8th International Conference on
Ubiquitous Information Management and Communication, ACM,
pp. 56, 2014.
[2] J.D. Haynes and G. Rees, “Decoding mental states from brain
activity in humans”, Nature Reviews Neuroscience, vol. 7,
pp. 523–534, July 2006.
[3] H. L. Niculescu, et al., "Discovery and validation of blood
biomarkers for suicidality", Molecular psychiatry, vol. 18,
no. 12, pp. 1249-1264, Aug. 2013.
[4] Siddiqi, Muhammad Hameed and Alam, Md Golam Rabiul
and Hong, Choong Seon and Khan, Adil Mehmood and
Choo, Hyunseung, “A Novel Maximum Entropy Markov
Model for Human Facial Expression Recognition,” PloS one,
vol. 11, no. 9, 2016.
[5] Ahmed A. A. Esmin; Roberto L. De Oliveira Jr., Stan Matwin,
“Hierarchical Classification Approach
to Emotion Recognition in Twitter,” 11th International
Conference on Machine Learning and Applications, vol. 2,
pp. 381-385, 2012.
[6] Sander Koelstra, Christian M¨uhl, Mohammad Soleymani,
Jong-Seok Lee, Ashkan Yazdani, Touradj Ebrahimi, Thierry
Pun, Anton Nijholt and Ioannis Patras, “DEAP: A Database
for Emotion Analysis Using Physiological Signals,” IEEE
Transaction on Affective Computing, vol. 3, no. 1, pp. 18-31,
2012.
[7] Md Golam Rabiul Alam,, Rim Haw, Sung Soo Kim, Md Abul
Kalam Azad, Sarder Fakhrul Abedin, and Choong Seon
Hong. "EM-psychiatry: An Ambient Intelligent System for
Psychiatric Emergency." IEEE Transactions on Industrial
Informatics, vol. PP, No. 99, 2016.
[8] Russell JA. “A circumplex model of affect”, Journal of
Personality and Social Psychology, vol. 39, No. 6, 1980.
[9] Md Golam Rabiul Alam, Sarder Fakhrul Abedin ,
Moshaddique Al Ameen and Choong Seon Hong, " Web of
Objects Based Ambient Assisted Living Framework for
Emergency Psychiatric State Prediction," Sensors, Vol. 16,
No. 9, September 2016.
[10] J. Schmidhuber, Deep learning in neural networks: An
overview, Neural networks, Elsevier, vol. 61, pp. 85-117,
2015.
2017년 한국소프트웨어종합학술대회 논문집
503