Cross-phase Emotion Recognition using Multiple Source Domain
Ke-Ming Ding
, Tsukasa Kimura
, Ken-ichi Fukui
and Masayuki Numao
Graduate School of Information Science and Technology, Osaka University, Japan
The Institute of Scientific and Industrial Research (ISIR), Osaka University, Japan
Electroencephalography (EEG), Emotion, Domain Adaptation, Generative Adversarial Network (GAN),
Cross Phase.
EEG signal, the brain wave, has been widely applied in detecting human emotion. Due to the human brain’s
complexity, the EEG pattern varies from different individuals, leading to low cross-subject classification per-
formance. What is more, even within the same subject, EEG data also shows diversity for the same reason.
Many researchers have conducted experiments to deal with the variance between subjects by transfer learning
or domain adaptation. However, most of them are still low-performance, especially when the new subject
does not share generality with training samples. In this study, we examined using cross-phase data instead
of cross-subject data because the discrepancy of different phase data should be smaller than that of different
subjects. Different phases represent data recorded multiple times from the same subject with the same stimuli.
Two neural networks are adopted to verify the effectiveness of the cross-phase domain adaptation. As a result,
experiments on the public EEG dataset showed approximation level accuracy compared to the state-of-the-art
method but much lower standard derivation. Moreover, multiple source domains promote accuracy in contrast
to one single domain. This study helps develop a more robust and high-performance real-time EEG system by
transferring knowledge from previous data phases.
Emotion plays an important role in cognitive science
to decode the human brain. Traditional methods in de-
tecting affective states include questionnaires or com-
munication with the subject directly while these in-
teractions intensely depend on individual perception.
However, results are steadily influenced by exter-
nal factors like environment, preference and tension.
In recent years, using psychological signals, espe-
cially the electroencephalography (EEG), eye tracks,
magnetoencephalography (MEG), is extensively em-
ployed to detect emotion state due to its objectivity
(Bos et al., 2006). Among all these signals, EEG is
mostly adopted because of its non-invasive and man-
ageable solution (Coan and Allen, 2004). EEG elec-
trodes placed at the scalp surface continuously record
signals and trace affective states simultaneously. Ma-
chine Learning technique is thus adopted in detecting
discriminative features and emotion classification de-
pending on the electrical signals. Traditional machine
learning algorithms include linear discriminant anal-
ysis (LDA), support vector machines (SVM) (Duan
et al., 2013). With the development of deep learning,
there are more and more deep neural networks em-
ployed in bio-signal processing as a feature extractor
or classifier (Kahou et al., 2016). These deep learn-
ing architectures show more robust and better per-
formance in many classification tasks. However, the
diversity of EEG patterns within individual subjects
limits the transferability of trained models between
them, that is, model trained will not be necessarily
profitable in a new subject (Jayaram et al., 2016). Al-
though in a within-subject experiment, advanced clas-
sifiers obtain quite high accuracy that over 90%, in a
subject-independent experiment, the same algorithm
can only get accuracy about 70% (Luo et al., 2018),
which further proves the high discrepancy across dif-
ferent subjects. Domain Adaptation (DA), is pro-
posed to deal with the problem, which tries to align
the distribution of different subjects to train a com-
mon classifier upon the new distribution (Margolis,
2011). For example, a DA algorithm named adaptive
subspace feature matching (ASFM) (Chai et al., 2017)
reduced dimension both on the source domain and tar-
get domain by PCA. Then the algorithm integrated
Ding, K., Kimura, T., Fukui, K. and Numao, M.
Cross-phase Emotion Recognition using Multiple Source Domain Adaptation.
DOI: 10.5220/0010200701500157
In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 150-157
ISBN: 978-989-758-490-9
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
both the marginal distributions with a unified trans-
formation function. As a result, both the marginal
and conditional discrepancies between the source and
target domain were reduced. Finally, logistic regres-
sion was applied to the new source subspace and a
classifier was adopted in the target domain. ASFM
shows significant improvement in accuracy compared
with a simple no-adaptation method, showing accu-
racy about 80%.
By now, many domain adaptation technologies
have been applied in subject-independent emotion
classification and achieve improvement in accuracy
(Jayaram et al., 2016; Zheng and Lu, 2016). They
want to take advantage of a large amount of data
from different subjects and extract common knowl-
edge that can apply in any new subject. However, we
need to note that the variance within different subjects
still exists and may lead to negative transfer learn-
ing——if the new subject is quite different from any
other training samples (Li et al., 2018). Furthermore,
collecting abundant training samples from different
subjects can be time-consuming and costly, even im-
possible in build real-time emotion recognition sys-
tem (Guger et al., 2000). To solve this problem, this
study utilized data recorded from the same subject in-
stead of data from different subjects so that the dis-
crepancy will be smaller because of the correlation
of emotion within the same subject. We used one of
the most-cited EEG emotion dataset SEED and sub-
jects in this dataset watched the same movie clip three
times. Instead of data from different individuals as
source domain, we choose data from the same per-
son but in different periods as the source. The term
phase is applied in reference to separate EEG sig-
nals while the subject was exposed to the same stim-
uli. Several unsupervised domain adaptation meth-
ods, domain-Adversarial Neural Network (DANN)
(Ganin et al., 2016) and Multiple Source Domain
Adaptation (MDAN) (Zhao et al., 2018) were oper-
ated and the results were compared with one state-
of-the-art (SOTA) approach. The classification ac-
curacy of MDAN was in approximation level with
the SOTA model and MDAN showed much lower
standard derivation. In addition, MDAN indicated
more potential in real-world application. Increasing
more source domains (collecting more data from the
same subject) helps promote classification. What is
more, we investigated what actually conclude nega-
tive transferring. Even exposure to the same movie,
the familiarity and likeness will change while only
similar affection enhances transfer learning. We ex-
plored the self-assessments when each subject expos-
ing to the stimuli and compared the difference rating
scores. The result further convinces that reaction of
disparate subjects will be much more variant than re-
action from one same subject, revealing the effective-
ness of our research.
2.1 SEED Dataset
This study was performed on the publicly available
EEG dataset for emotion recognition. SEED dataset
is proposed by Shanghai JiaoTong University (SJTU)
( The SEED dataset
consists of 15 participants, each of whom needs to
watch 15 Chinese movie clips to induce three kinds
of emotions: positive, neutral, and negative. All of
them are native Chinese with 7 males and 8 females.
The movie clips are well selected as the stimuli in the
experiments. All movies need to be understandable
and neither too long nor too short to induce enough
emotional meanings. The EEG signals were recorded
by an ESI NeuroScan System with a 62 electrode cap
at a sampling rate of 1000 Hz. After each exploration
of a movie clip, the subject needs to rate their emo-
tion. Each film clip lasts about 4 minutes and there
is a 5-s hint before each clip and 46-s for a question-
naire. Questions are as follows for self-assessment:
1) what they had felt in response to viewing the film
clip; 2) have they watched this movie before; 3) have
they understood the film clip. The procedure of each
trail is shown in Figure 1.
Figure 1: Protocal of each trail in the experiment.
Compared with many other EEG emotion datasets
like DEAP (Koelstra et al., 2011), which only
recorded data once when subjects were exposed to
the stimuli, the SEED dataset has a special term of
session. Each session includes data of all the subjects
watching all the 15 film clips once. There were 3 ses-
sions of data collected during the experiment. The
difference between these sessions is the external en-
vironment like the time of the experiment. The mood
of each subject and familiarity of each movie must
be different since we can’t strictly control the internal
variables of each participant. However, the stimuli
and subjects are the same, which guarantees the la-
tent correlation of each session data. One of the main
purposes of this study is to discover the latent corre-
Cross-phase Emotion Recognition using Multiple Source Domain Adaptation
lation of EEG recording from the same person when
exposing to the same stimuli.
2.2 Preprocess and Feature Extraction
The raw data was down-sampled with a frequency of
200 Hz and artifacts are removed manually. EEG sig-
nals that were seriously contaminated by EMG and
EOG were checked visually and removed manually.
Eye tracks (EOG) were also recorded to mark arti-
facts and removed later. A band-pass filter between
0.3 and 50 Hz was utilized to remove noise upon
the EEG signals. Only EEG segments in duration
of each film are selected. A 1s Hanning window
without overlapping is operated on the original EEG
data and finally, 3300 clean epochs are obtained in
each channel for each experiment. Instead of raw
data, frequency domain data, which calculated by a
512-point short-time Fourier transform in each time
window is applied as input features. Various feature
types are provided in SEED dataset like power spec-
tral density (PSD), differential asymmetry (DASM),
rational asymmetry (RASM), asymmetry (ASM), and
differential entropy (DE). According to some studies
(Zheng and Lu, 2015; Duan et al., 2013), a simple
but discriminative feature named DE feature shows
good performance in emotion recognition compared
to other features. DE feature is defined as :
h(X) =
where x obeys the Gaussian Distribution N
µ, σ
Obviously, original EEG signal doesn’t follow the
certain distribution. However, studies (Duan et al.,
2013; Zheng and Lu, 2015) proved that the probabil-
ity that certain sub-band signals meet Gaussian distri-
bution is larger than 90 percent. It has been proved in
that for a fixed EEG sequence, DE is equivalent to the
logarithm average energy in a certain frequency band.
After band-pass filtering in certain bands (delta: 1-3
Hz, theta: 4-7 Hz, alpha: 8-13 Hz, beta: 14-30 Hz,
gamma: 31-50 Hz), Therefore, DE feature is calcu-
lated as the input feature (Duan et al., 2013) following
the equation. Here, all 62 channels are used so the fi-
nal input shape is 62 channels x 5 bands. One session
of the total 15 subjects has a total of 27430 samples
of data so the shape of all 3 sessions is (27430, 62, 5).
Table 1: Part of self-assessment from different subjects.
subject ID session
score of each film clip
1 2 3 4 5 6
1 5 5 5 4 5 3
2 4 5 5 4 4 4
3 4 5 4 4 5 5
1 1 3 3 1 4 3
2 2 3 2 2 4 3
3 2 4 2 2 4 3
1 4 4 5 5 3 4
2 4 4 5 5 4 4
3 3 4 4 5 4 4
1 5 3 4 4 3 4
2 5 3 5 5 4 4
3 5 5 3 5 4 4
1 4 3 5 4 5 5
2 5 5 5 4 5 5
3 4 4 4 3 4 5
2.3 Emotion Labeling
Different from many other EEG emotion dataset,
SEED notifies the label of each film clip as the ground
truth, considering the selected film clip induced target
emotion successfully. This labeling approach works
better than traditional self-assessment from subjects.
Participants do not have a good understanding of dis-
crete indicators like 10-scale variance, who are not
experts in psychology. The point they chose only
shows the tendency of their mood contrary to a certain
threshold rather than the correct tension to the stimuli.
The SEED dataset provides another table that records
the self-assessment after watching each film clip. We
listed part of the rating score from different subjects
and the reaction in each session in Table 1. Each score
means variance of emotion in a range from 0 to 5. The
table indicates that even when the subject watching
the same movie, the rating score changes, but within
a relatively stable range. This result suggests the dis-
crepancy of session data exists even within the same
subject and using movie labels instead of assessment
labels helps build a more robust model.
3.1 Domain Adaptation
Domain Adaptation is a subset of transfer learning,
striving to align data distribution from different do-
mains. It is widely applied in the field of image
classification aiming to get high performance in the
label-lack domain using label-rich domain data. In
BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing
reverse to traditional transfer learning methods like
fine-tuning, domain adaptation is more likely to be
unsupervised machine learning (Pan and Yang, 2009).
Assume that we have two datasets which are in dif-
ferent distribution.
, Y
represents the labeled
source dataset and {X
} represents the non-label tar-
get dataset, where m and n represent the number of
sample data in source and target domains respectively.
We want to find a feature mapping that transforms the
source X
and target data X
into a common subspace
where X
and X
have the same data distribution, so
that we can achieve good performance in both source
domain and target domain.
3.2 Domain-Adversarial Neural
Network (DANN)
Overview of GAN: Generative Adversarial Network
(GAN) was proposed in 2014 (Goodfellow et al.,
2014). Till now, GAN has been extensively applied in
many fields like data segmentation, auto image gen-
eration, and domain adaptation. In a traditional GAN,
two main components are optimized: the discrimina-
tor D and the generator G, both of which are com-
posed by neural networks. G and D have different loss
functions and the training process minimizes the loss
separately, playing a minimax game. The discrimina-
tor tries to classify whether the generated data is true
opposed to the original input data while the generator
tries to fool D by producing generated samples that
are similar to the real data. The final objective func-
tion is defined mathematically as follow:
log(1 D(G(z)))
Since GAN shows remarkable capability in gener-
ating controllable data distribution, many researches
implement domain adaptation with GAN in many
fields. Domain Adversarial Neural Network (Ganin
et al., 2016) is such a GAN-based network that trains
in an unsupervised approach to get proper feature
mapping space. Figure 2 shows the architecture
of DANN. There are three main components in the
DANN, the feature extractor, the label predictor, and
the domain classifier. The feature extractor consists
of several layers of CNN, which do feature extraction
upon both source and target domain data. To normal-
ize the feature mapping after each CNN layer, there’s
a Batch-Normalization layer which makes the fea-
ture mapping obey a certain distribution. After sev-
eral CNN+BN layer combinations, there’s an average
pooling layer to reduce the feature shape. As for the
label predictor, it’s made by some full connection lay-
ers and final softmax classifier. Only source domain
data will be sent into the label predictor since no-label
target domain data can be optimized by the classifi-
cation loss. The domain classifier is quite similar to
the label predictor while the final softmax classifier
does binary classification, taking human-made labels
where data from the source has a label of 1 and data
from the target has a label of 0. The domain classifier
attempts to predict whether the data comes from the
source or the target. Besides, a gradient reversal layer
is set between the feature extractor and domain clas-
sifier to train the whole network in a backpropagation
Figure 2: Architecture of DANN.
Assume that {X
, Y
} represents data from the
source and {X
} represents data from the target. The
feature extractor processes both two domain data and
yields feature mapping, the {X
} and {X
}. The
label predictor predicts emotion labels by optimiz-
ing a cross-entropy loss while the domain classifier
makes source/target prediction using human-made la-
bels. The gradient reversal layer represents a negative
constant and the total loss function will be:
E (θ
, θ
, θ
) =
), y
(R (G
), d
Training in such an adversarial approach by min-
imizing the label loss and maximizing the domain
loss, a high-performance label predictor and low-
performance domain classifier were obtained, mean-
ing the source data was well learned and the distribu-
tions were well aligned.
3.3 Multiple Domain Adaptation
Network (MDAN)
MDAN is an extension of the DANN, which aligns
numerous data distributions in an adversarial training
process (Zhao et al., 2018). The difference is that
MDAN uses two source domains with two domain
classifiers. Data from two different source domains is
notated by {X
, X
} with labels {Y
, Y
} and target
domain data is X
without label. There are two main
Cross-phase Emotion Recognition using Multiple Source Domain Adaptation
purposes of this MDAN: (1) distinguishable enough
for label predictor in all source domain data. (2) in-
distinguishable enough for the two domain classifiers
to separate data between source domains and target
domain. The architecture of MDAN is in Figure 3.
Similar to the DANN, the feature extractor con-
sists of several layers of CNN and BN layers. It ex-
tracts low-level features from the original data space
in both source and target domains. Optimization func-
tion in task label predictor is:
, θ
And the objective function in two domain classifiers
is quite similar:
, θ
) +
, θ
The reason why we use this neural network is that:
the discrepancy between EEG data in one single sub-
ject cross-phase exists, making it difficult to build a
real-time emotion recognition system——the model
trained by formerly collected data can’t be adapted
directly in present collected data. However, in con-
trast to cross-subject data, we note that cross-phase
data has stronger correlation which can be strength-
ened by MDAN.
Figure 3: Architecture of MDAN.
In this section, the effectiveness of DANN is de-
scribed as well as the no-adaptation method and sin-
gle domain adaptation method. All the experiments
were performed with SEED dataset. There are in to-
tal 3 sessions data as independent domains for eval-
uation. This helped us investigate whether these DA
methods overcome the non-stationarity of EEG sig-
nals. One session refers to all subjects exposed to
stimuli once and 3 sessions were acquired by repe-
tition at one week intervals. One significant contribu-
tion of the SEED dataset is that each subject repeated
the same experiments three times. We have a way
to further inspect the spatial-relationship of the affec-
tive state in each individual participant. Also, since
Table 2: Best accuracy (%) of MDAN and DANN.
1+2/3 72.40 85.84 81.32 74.80
1+3/2 70.72 84.89 82.09 74.32
2+3/1 73.01 83.64 79.44 81.76
SEED is a widely used dataset and some session-
independent research has been done already, we made
comparison with these methods.
4.1 One Source vs Two Source Domain
To build a real-time emotion recognition system, we
hope to have good classification accuracy in the tar-
get domain which has no notation. However, since
the discrepancy of different phase data exists, just
concatenating different phase data together doesn’t
make sense, even leading to negative transfer learn-
ing. The DANN takes one session as source do-
mains so we choose the best accuracy when trying
either of two source domains. In contrast, MDAN
utilizes the both two source domains. The classifi-
cation accuracy is shown in Table 2. We can see
a naive combination of different session data didn’t
have good performance and even one-source domain
adaptation (DANN) outperformed a lot. MDAN per-
formed much better than DANN using combined data.
The result showed GAN-based domain adaptation has
a good effect in aligning data distribution from differ-
ent phase data and more data shows more improve-
ment in classification accuracy.
Note that in DANN, choosing different source do-
main made influence to the accuracy to some degree.
The reason may relay on change of affective states
since session 1 refers to the first time subjects watch-
ing the movies and session 3 refers to the third time.
Session 2 acts as a intermediary which has closer cor-
relation to session 1 and 3. If we go further into the
self-assessment from subjects and calculate the MSE
error between every two sessions, we can have a bet-
ter understanding of the cross-phase relationship. Re-
sults in Table 3 accounts for why DANN in 1 to 3
session adaptation performs poor. The assessment be-
tween these 2 sessions are much higher than others,
showing quite massive transformation happened dur-
ing the first experiment and the third.
The rating point of one typical subject in whole
3 sessions is plotted in Figure 4. In extreme case like
movie 12, 5-scale rating can be thoroughly distinctive,
as large as 3 points. Even the ground truth of the film
is neural.
BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing
Table 3: MSE of self-assessment between every 2 sessions.
Combination MSE
1 and 2 5.633
2 and 3 8.183
1 and 3 5.917
Figure 4: Rating score of subject 14.
We trained the model with a mini-batch of 64 us-
ing optimizer Adam. The total training epoch was
1000 until the accuracy curve was converged. The
initial learning rate was 0.001 and decayed with run-
ning epoch. The accuracy curve of all three domains
is plotted in Figure 5. As often the case when training
with GAN, the domain accuracy fluctuates with high
amplitude and will never be converged. The train-
ing accuracy in both source domains goes to 100%
rapidly in contract with the target domain accuracy.
Figure 5: Training Accuracy curve of MDAN (prediction
accuracy in the left and domain accuracy in the right).
Furthermore, we can have a look at the data distri-
bution after dimension reduction algorithm of t-SNE
(Maaten and Hinton, 2008). The original 310 dimen-
sion feature (62 channels x 5 bands) was decomposed
into two dimension space. In Figure 6, original fea-
ture space are made by several clusters which are dis-
tinguishable. These clusters should be data from dif-
ferent subjects and different colors refer to the 3 ses-
sions. The discrepancy across subjects is far more
considerable than that across sessions, which makes
it harder to transfer. In the right figure, distribution of
entire 3 session data mix together instead of clustering
separately, showing the MDAN successfully aligned
the data distribution.
Figure 6: 2-Dimension Data Distribution of MDAN after
t-SNE (red,blue,green colors refer to three session data).
Other session-independent methods are also applied
with the SEED dataset, however, most of them can
only get medium performance considering the vari-
ance of EEG signal. Instead of analysing the cor-
relation of phase data within the same subject, they
regard each session data as separate dataset. The pur-
pose of all these methods is to get an aligned sub-
space between source and target domains. There are
two main categories in related works, one is using ad-
versarial training to get an aligned distribution, rep-
resentative works like the DANN and MDAN. The
other is a dimension-reduction based method, hop-
ing to get a no-variance low dimension feature from
original high dimension feature such as ASFM. Since
we hope to evaluate the transferability between dif-
ferent phases and the enhancement of multiple source
domain data, DANN and MDAN can be desirable
methods. One state-of-the-art method and some well-
known machine learning techniques are introduced
here as baselines to the MDAN.
5.1 Adaptive Subspace Feature
Matching (ASFM)
In this study (Chai et al., 2017), Principal Component
Analysis (PCA) was chosen to do subspace align-
ment. Among the full space, d largest eigenvalues
were selected by PCA and worked as the bases of both
source and target subspace, Zs and Zt. Compared to
large feature space domain adaptation which requires
a large amount of computation, dimension reduction
shows better performance in online implementations.
The ASFM tries to find the best feature map from
source space Zs to target space Zt in linear transfor-
mation. The main process of ASFM is in Figure 8.
Cross-phase Emotion Recognition using Multiple Source Domain Adaptation
Figure 7: Architecture of ASFM.
5.2 Comparison with Other Methods
There are many other domain adaptation methods
to deal with the EEG emotion recognition problem.
Here, we compared with the state-of-the-art method,
the Adaptive Subspace Feature Matching (ASFM).
ASFM tends to be a traditional machine learning al-
gorithm that Linear Regression works as the base
classifier. Three famous machine learning methods,
the Support Vector Machine (SVM), Linear Regres-
sion (LR) and Auto Encoder (AE) are adopted as
baseline methods. Table 4 presents mean accuracy
and standard deviation of all these methods (Chai
et al., 2017). No adaptation methods like SVM shows
average accuracy about 70%, higher than channel
level (33%) but in a high standard deviation. ASFM
achieves best mean accuracy and relatively small stan-
dard deviation in most of the cases. However, none of
them has utilized the advantage of multiple source do-
mains since all these methods don’t have a way to deal
with the multiple invariances between several source-
target pairs. The MDAN we used here addressed
this problem and gave a hint that more source do-
mains help improve the performance. Moreover, both
DANN and MDAN are deep neural networks, which
are sensitive to hyper-parameter tuning and training
epochs. To get a fair result, we repeated the same ex-
periments five times to calculate the mean and stan-
dard deviation (SD). The mean accuracy and standard
deviation of MDAN is 83.40% and 0.96%. The SD is
much smaller than that in ASFM, demonstrating sta-
ble performance using MDAN and more potential in
real-time recognition task.
Figure 8: Results of other DA methods(Chai et al., 2017).
The results presented above show the MDAN was a
promising technique in transferring previous phase
Table 4: Average accuracy and standard deviations (%) of
other DA methods. (* mark means best accuracy).
1 2 3 Average
Figure 9: Comparison with SOTA method.
domain knowledge to predict in new phase domain
data. There are two main contributions of this work.
1). We adopted the MDAN network which is an
unsupervised machine learning method to align data
distribution among different domains and get ap-
proximation level accuracy in comparison with the
state-of-the-art method. 2). Instead of traditional
subject-independent evaluation, experiments between
sessions was performed. Experiment results reveal
that transferring knowledge from different phases of
data in one same subject helps build a more robust
and high-performance real-time EEG system. For fu-
ture work, we plan to explore what actually affect
the emotion state in each subject when repeatedly ex-
posed to a same stimuli. SEED dataset only consisted
data from 15 subjects, which is far more than general-
ization. More subject data is required to validate the
effectiveness of MDAN and further feature extraction
method should take into consideration as general. As
the case in some subject, the rating score switched a
lot even when the film was well selected to induce
specific emotion. To the end, since the main focus is
on cross-phase domain adaptation, further data from
same subject is also necessary to certify that more
phase data brings about improvement in accuracy.
BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing
Bos, D. O. et al. (2006). Eeg-based emotion recognition.
The Influence of Visual and Auditory Stimuli, 56(3):1–
Chai, X., Wang, Q., Zhao, Y., Li, Y., Liu, D., Liu, X.,
and Bai, O. (2017). A fast, efficient domain adap-
tation technique for cross-domain electroencephalog-
raphy (eeg)-based emotion recognition. Sensors,
Coan, J. A. and Allen, J. J. (2004). Frontal eeg asymmetry
as a moderator and mediator of emotion. Biological
psychology, 67(1-2):7–50.
Duan, R.-N., Zhu, J.-Y., and Lu, B.-L. (2013). Differential
entropy feature for eeg-based emotion classification.
In 2013 6th International IEEE/EMBS Conference on
Neural Engineering (NER), pages 81–84. IEEE.
Ganin, Y., Ustinova, E., Ajakan, H., Germain, P.,
Larochelle, H., Laviolette, F., Marchand, M., and
Lempitsky, V. (2016). Domain-adversarial training of
neural networks. The Journal of Machine Learning
Research, 17(1):2096–2030.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Ben-
gio, Y. (2014). Generative adversarial nets. In
Advances in neural information processing systems,
pages 2672–2680.
Guger, C., Ramoser, H., and Pfurtscheller, G. (2000). Real-
time eeg analysis with subject-specific spatial patterns
for a brain-computer interface (bci). IEEE transac-
tions on rehabilitation engineering, 8(4):447–456.
Jayaram, V., Alamgir, M., Altun, Y., Scholkopf, B., and
Grosse-Wentrup, M. (2016). Transfer learning in
brain-computer interfaces. IEEE Computational In-
telligence Magazine, 11(1):20–31.
Kahou, S. E., Bouthillier, X., Lamblin, P., Gulcehre,
C., Michalski, V., Konda, K., Jean, S., Froumenty,
P., Dauphin, Y., Boulanger-Lewandowski, N., et al.
(2016). Emonets: Multimodal deep learning ap-
proaches for emotion recognition in video. Journal
on Multimodal User Interfaces, 10(2):99–111.
Koelstra, S., Muhl, C., Soleymani, M., Lee, J.-S., Yazdani,
A., Ebrahimi, T., Pun, T., Nijholt, A., and Patras, I.
(2011). Deap: A database for emotion analysis; using
physiological signals. IEEE transactions on affective
computing, 3(1):18–31.
Li, X., Song, D., Zhang, P., Zhang, Y., Hou, Y., and Hu, B.
(2018). Exploring eeg features in cross-subject emo-
tion recognition. Frontiers in neuroscience, 12:162.
Luo, Y., Zhang, S.-Y., Zheng, W.-L., and Lu, B.-L.
(2018). Wgan domain adaptation for eeg-based emo-
tion recognition. In International Conference on Neu-
ral Information Processing, pages 275–286. Springer.
Maaten, L. v. d. and Hinton, G. (2008). Visualizing data
using t-sne. Journal of machine learning research,
Margolis, A. (2011). A literature review of domain adapta-
tion with unlabeled data. Tec. Report, pages 1–42.
Pan, S. J. and Yang, Q. (2009). A survey on transfer learn-
ing. IEEE Transactions on knowledge and data engi-
neering, 22(10):1345–1359.
Zhao, H., Zhang, S., Wu, G., Gordon, G. J., et al. (2018).
Multiple source domain adaptation with adversarial
Zheng, W.-L. and Lu, B.-L. (2015). Investigating criti-
cal frequency bands and channels for eeg-based emo-
tion recognition with deep neural networks. IEEE
Transactions on Autonomous Mental Development,
Zheng, W.-L. and Lu, B.-L. (2016). Personalizing eeg-
based affective models with transfer learning. In Pro-
ceedings of the Twenty-Fifth International Joint Con-
ference on Artificial Intelligence, pages 2732–2738.
Cross-phase Emotion Recognition using Multiple Source Domain Adaptation