Effectiveness of Cross-Model Learning Through View-Model

Ensemble on Detection of Spatiotemporal EEG Patterns

Ömer Muhammet Soysal

, Iphy Emeka Kelvin

and Muhammed Esad Oztemel

Computer Science, Southeastern Louisiana University, Hammond, U.S.A.

Division of Electrical and Computer Engineering, Louisiana State University, Baton Rouge, U.S.A.

Keywords: Autoencoder, Biometric, Brain, Convolutional Neural Network, Electroencephalograph, Spatiotemporal,

Transfer Learning.

Abstract: Understanding the neural dynamics of human intelligence is one of the top research topics over the decades.

Advances in the computational technologies elevated the level of solving the complex problems by means of

the computational neuroscience approaches. The patterns extracted from neural responses can be utilized as

a biometric for authentication. In this study, we aim to explore cross-model transfer learning approach for

extraction of distinct features from Electroencephalography (EEG) neural signals. The discriminative features

generated by the deep convolutional neural network and the autoencoder machine learning models. In addition,

a 3D spatiotemporal View-matrix is proposed to search distinct patterns over multiple EEG channels, time,

and window segments. We proposed a View-model approach to obtain intermediate predictions. At the final

stage, these intermediate scores are ensembled through a majority-voting scheme to reach the final decision.

The initial results show that the proposed cross-model learning approach can outperform the regular

classification-based approaches.

1 INTRODUCTION

Machine learning has been utilized in different

electroencephalography- related research including

brain computer interface (Aggarwal, 2021), diagnosis

of neurological disorders (Oh, 2020), human

computer interaction (Zhao, 2020), development of

authentication systems (Fidas and Lyras, 2023) and

many others (Khosla, 2020). As non-invasively

collected data, EEG recordings exhibits both spatial

and temporal features for comprehensive analysis of

human brain characteristics.

Intra-subject characteristics of EEG signals

demonstrate similar patterns extracted over various

trials while they differ significantly among the

subjects (Mueller, 2013). This distinctiveness

property allows EEG patterns to be utilized as a

biometric for personal authentication. Various

advantages of identification based on brain signals

have been emphasized compared to traditional

personal verification methods (Bidgoly, 2020). For

example, fingerprint, retinal scan, voice recognition

and facial recognition systems may have

vulnerabilities in terms of data security and deception

attempts against these systems (Bharadwaj, 2014).

However, brain signals can provide a more secure

personal verification method against such threats

(Riera, 2007).

The fusion of multi-view predictions can improve

classification performance (Xu, 2013). Among

various fusion strategies, (Kuncheva, 2014) and

(Atrey, 2010) showed that the majority-voting

scheme performed better than single-view decision

making.

In this study, we explored effectiveness of cross-

model based learners that generate feature patterns

from proposed spatiotemporal View-matrix utilizing

a multi-view ensemble classifier system. The

proposed approach is unique in terms of introducing

1) a cross-model transfer learning framework that

employs the DCNN and the AE with widely used

regular classifiers and 2) testing the performance of

the proposed system using cross-session datasets. The

rest of the report is organized as follows: The method

section starts with describing the data acquisition and

preparation procedure. The section flows with

presenting the proposed View-Matrix data structure,

View-Model, and ensemble of these models. The

result section is discussing the effectiveness of the

proposed framework and hyperparameter scheme.

942

Soysal, Ö. M., Kelvin, I. E. and Oztemel, M. E.

Effectiveness of Cross-Model Learning Through View-Model Ensemble on Detection of Spatiotemporal EEG Patterns.

DOI: 10.5220/0013265300003912

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages

942-949

ISBN: 978-989-758-728-3; ISSN: 2184-4321

The conclusion summarizes outcomes and points to

limitations that can be improved.

1.1 Previous Work

Machine learning techniques are widely used for

verification of individuals based on EEG patterns.

Fidas et al. discussed the role of machine learning

techniques in personal verification applications. In

summary, wavelet transform, power spectral density,

autoregressive modelling and fast Fourier transform

are the main techniques used for feature extraction.

Support vector machine, hidden Markov models,

multilayer perceptron, recurrent neural networks

(RNN) and convolutional neural networks (CNN)

were used in the classification of the data obtained

from these features.

Autoencoders have been studied for different

purposes in analysis of brain signals (Weng, 2024).

As an example, AE mechanism can be used to remove

eye blink artifacts from EEG signals (Acharjee,

2024), to identify sleep stages (Dutt, 2022).

(Abdelhameed, 2018) utilized an AE network to

predict epileptic seizures. (Bandana, 2024) employed

a spatial AE network for personal verification. Latent

features obtained from the AE network trained a CNN

model. Ari et al. pointed out that AEs provide an ideal

solution for artificial data generation to increase the

amount of training data (Ari, 2022). Tian et al.

operated two encoders simultaneously (Tian, 2023).

Zhou and Wang utilize spatiotemporal AE, with

adaptive diffusion method, to obtain high resolution

EEG data from low-resolution data (Zhou, 2024).

Yao and Motani stated that the vital signs of the

patients contain both temporal and spatial

information. Therefore, they proposed a hybrid

learning mechanism for classification purposes. Their

system first extracts spatial features, and then

temporal patterns to determine if an individual is an

alcoholic or not. Support vector machine, gradient

boosting, random forest and decision tree algorithms

were applied for classification. Among these

classification techniques, SVM achieved the most

successful results (Yao, 2018).

On the other hand, multi-view fusion models

provide improved performance over single view-

based classification. Mane et al. examined multi-view

features obtained from different frequency bands to

train a CNN (Mane, 2020). Spyrou et al. applied

multi-view tensor factorization for detection of

epilepsy by means of a linear regression method

(Spyrou, 2015). (Gao, 2022) compared the multi-

view and single-view classification for emotion

recognition; it was found that the multi-view

classification is superior to single-view. (Emanet,

2024) employed multi-view hierarchical learning

model with 3D-CNN for classification of a stimulus

type. Jia et al. aimed to classify sleep stages utilizing

spatial-temporal graph convolutional network

through multiple views that are consisted of

functional connections and distance-based

connections (Jia, 2021).

Transfer learning techniques have been

successfully used in various EEG-related studies such

as motor imagery and evoked potential applications

(Wu, 2020). Waytowich et al. focused on

unsupervised spectral transfer learning and geometry-

based knowledge training for brain-computer

interface study examining subject independence

(Waytowich, 2016). Qi et al. used inter-subject

transfer learning to reduce the calibration time. A

small number of epochs for target subject is taken as

references and the Riemann distance metric was

calculated and applied to the most similar target

subject (Qi, 2018). Transfer learning can be applied

across devices as well as across subjects. Wu et al.

investigates how to improve the performance of

brain-computer interfaces (BCIs) by reducing the

amount of time needed to calibrate them for use with

different EEG headsets. The authors propose a new

method called active weighted adaptation

regularization (AwAR), which combines transfer

learning and active learning to facilitate the

calibration process. AwAR leverages data from

previously used EEG headsets to train a classifier for

a new headset, selecting only the most informative

data points for labelling. This significantly reduces

the amount of data required for calibration, ultimately

making BCI technology more user-friendly and

accessible (Wu, 2016). Additionally, Cimtay et al.

used a previously trained CNN model based on

Inception-ResNet in emotion recognition systems by

transferring its weights between subjects and datasets

(Cimtay, 2020).

2 METHOD

In this section, we describe the framework for

extraction of distinct patterns from spatio-temporal

EEG neural responses. The framework is composed

of the following main modules: 1) Preprocessing, 2)

feature extraction, 3) generating view-models, 4)

building fusion-model as illustrated in Figure 1.

Preprocessing is responsible for filtering artifacts

from the raw signal such as mean-line harmonics,

extraction of the spectrum of interest. The Laplacian

of Gaussian (LoG) is used as a signal conditioning

Effectiveness of Cross-Model Learning Through View-Model Ensemble on Detection of Spatiotemporal EEG Patterns

943

operator to enhance the signal. The View-Matrix

Generator formats the original 2D (channel, time)

data into a 3D spatiotemporal matrix. The View-

models are composed of the base learners, deep

convolutional neural network (DCNN) and AE, and

four regular classifiers namely k-nearest neighbours

(KNN), random forest (RF), support vector machines

(SVM), and artificial neural network to identify

participants. At the final stage, the fusion unit

ensembles View-model predictions to reach the final

decision.

Figure 1: Workflow.

The proposed method has been tested on EEG

data, which is composed of 7 subjects in 2 sessions,

10 days apart. We utilized the mBrain Smarting PRO

amplifier equipped with a 24-channel head cap. The

electrode locations on the head cap were designed

according to the 10-20 system. The amplifier was

configured at a sampling frequency of 500 Hz. We

designed and implemented several protocols using

the Presentation software: 1) Baseline, 2) inner voice-

audio, 3) shape-trace-audio, and 4) motion. Each

protocol is repeated for a total of 10 trials. In this

study, we presented results for the stimuli associated

with the resting state while eyes were open.

The EEG signal undergoes initial filtering with a

notch filter during the preprocessing stage. Next, a

band-pass filter is applied to focus on the 0.5 – 32 Hz

spectrum. The LoG operator described in (Oztemel,

2024) is then applied to enhance the signal. We

focused on trials with time duration of 0.5 seconds for

several stimuli. The EEG signal of a length L is

partitioned into different k segments of a length 𝑊 =

𝐿/(𝑝 𝑘)

with an overlap p.

2.2 3D Spatiotemporal Views

The proposed 3D Spatiotemporal View, named View-

Matrix, presented in Figure 3 combines the neural

dynamics over time for all EEG channels together

with cross-segment interactions. The neural activity

patterns are extracted from three views. View-1 is

composed of stacking the (channel, time) frames over

window segments. Similarly, View-2 enables

extraction of patterns in the stack of channel-window

frames over several periods and View-3 provides a

perspective from (window, time) frames across the

stack of channels.

Figure 2: 3D Spatiotemporal Views.

The descriptive feature patterns are generated

through a cross-model learning strategy. We explored

the effectiveness of the cross-model based transfer

learning over the regular classifiers (RC) ANN, KNN,

SVM, and RF. We utilized the DCNN and the AE as

base-learners. A View-model is generated by

employing an RC or combination of a base-learner

with an RC. Each View-model is constructed using its

designated View-matrix. When the training is

completed, the FC unit is dropped from the DCNN

model. Similarly, the decoder unit is discarded from

the AE model. The output of these models is utilized

to generate features passing through the flattening

unit F to train the regular classifiers. Figure 4

illustrates the training process flow. In the feature

extraction stage, a transfer-learning network model

generates features from the spatio-temporal set of

signals, named View-matrix. The fusion-model

combines predictions from multiple views to reach

the final decision.

2.3 Ensemble of View-Models

In the ensemble of intermediate predictions, we

utilized the idea of a voting classifier that produces

the final prediction from multiple opinions by

majority vote, i.e., the class with the highest

probability of being predicted by each classifier. The

fusion module yields the most frequently voted class

label together with the corresponding prediction

score. The mean prediction score is calculated when

more than one View-model predicts the same class.

When all three View-models disagree with each

other, a simple random selection determines the final

decision. Figure 5 illustrates the View-Model fusion

process.

View-1

(Channel, Time) → Window

Time

Channel

Time

View-2

(Channel, Window) → Time

View-3

(Window, Time) → Channel

Raw

Data

(Channel, Time)

View Matrix

Preproc

essing

View Matrix

Generator

Base-

Learner

View-Models

Fusion

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

944

Figure 3: Training Process of View-Models.

Figure 4: View-Model Fusion Process.

3 RESULTS

In this section, we present our findings on the

effectiveness of the proposed framework. Two

performance measures, the accuracy (ACC) and the

area under the curve (AUC), were utilized for

evaluation of the proposed framework. We compared

the regular learning models with the cross-model

learning networks, DCNN and AE. We employed a

5-fold cross-validation strategy to measure the

stability of the proposed framework against the

uncertainty of the data distribution. At each fold, we

split the data into 80% for training and 20% for

validation. For the assessment of permanence, the

session-1 EEG recordings were utilized to generate

the models, and the session-2 recordings were used

for testing. Hyperparameter tuning was conducted at

each fold. The duration of a segment of Interest (SoI)

was 0.5 seconds. The EEG amplifier operated at a

sampling frequency of 500 Hz. The session-1 and

session-2 included 514 SoIs, making 1028 SoIs in

total. The size of the View-matrix per subject was

241632 (channel, time, window).

3.1 Effectiveness of Cross-Model Learning

Figure 6 presents the effect of each learning scheme

for extraction of distinct patterns from the 3D View-

matrix. The regular classifiers ANN, KNN, and SVM

trained by the features directly flattening of a View-

Matrix performed poorly compared to the RF as

Figure 6a shows. In addition, the RF classifier did not

show a stable performance as its distribution was

quite wide.

On the other hand, the DCNN and AE-based

cross-learner models outperformed the regular

classifiers as shown in Figure 6b and Figure 6c. The

distribution of the average prediction scores elevated

significantly. It should be noted that the predictions’

stability requires attention to improve the proposed

approach.

The analysis of Figure 7 clearly shows that the

proposed cross-model approach significantly

improved the learning performance. Overall, the

DCNN base learner showed slightly higher prediction

scores on average than the AE’s predictions. As a

remark, there is room to conduct research on the

stability of the base learners.

3.2 Identifiability of Individuals

In this study, we present outcomes from our in-house

dataset of EEG recordings from 7 individuals. It is

expected that individuals can be distinguished from

one another due to the unique characteristics of their

brain’s anatomical and functional differences. In

Figure 8, we presented AUC performance values of

the classification algorithms for each subject. The

performances of the KNN and SVM models provided

very similar results. However, the random forest and

ANN models showed performance improvement in

some cases, depending on the utilized learning

approach. These findings show that the cross-learner

model with DCNN and RF combination can be more

successful for certain subjects.

3.3 Comparison with the State of the

Art

Arnau pointed out a common mistake in EEG-based

biometric studies (Arnau, 2021). Surprisingly, few

studies have focused on the effects of time-dependent

changes in brain signals. In most studies, systems

developed for high-accuracy detection of subjects

were typically trained and tested on data collected in

the same session. Alternatively, data collected from

different sessions were combined; and then split into

learning and testing datasets. As a result, their

performance scores were reported as high. Being

aware of this situation, some studies performed the

learning and testing phases using data collected from

completely different sessions. Nakamura et al.

analysed two different scenarios in their study

focusing on this issue. In the first scenario, learning

and testing data were collected from the same session,

while in the second scenario, data were obtained from

DCNN layers

HP Tuning

View

Encoder

Decoder

View-Model

(Channel, Time)

View-Model

(Channel, Window)

View-Model

(Window, Time)

Fusion

Model

Effectiveness of Cross-Model Learning Through View-Model Ensemble on Detection of Spatiotemporal EEG Patterns

945

different sessions. In the second scenario, the time

difference between the sessions varied from 5 to 15

days (Nakamura, 2017). As Arnau emphasized, it has

been proven that performance was higher when data

from the same session were used. It is worths

mentioning that one of the recent rare studies

(Plucińska, 2023), a spectral-based biometric

verification experiment, resulted in 75 to 96% ACC

depending on whether the cross-session data were

used for training and testing. A simple ANN classifier

was utilized to extract distinct features.

In this study, we used data from one session to

train the models and data from another session for

testing. The data collection sessions were completed

with a 10-day interval. Therefore, this study provides

one of the unique reports in the literature in terms of

isolating training and testing datasets. To the best of

our knowledge, this study is the first to propose a

multi-view cross-session framework for EEG-based

authentication utilizing a cross-session test dataset.

3.4 Hyperparameter Tuning

Table 1 and Table 2 provide insight into the

hyperparameters of each model. We utilized Bayesian

optimization in Keras Tuner to identify the best

parameters for our models. The KNN’s neighbour

parameter reduced from 11 to 4 (3) when trained by

the base-learner AE (DCNN). The SVM changed its

kernel type from polynomial to linear when used with

both base learners. The RF’s max_depth parameter

dropped from 14 to 5 and 11 when trained by the

DCNN and AE, respectively. The number of layers

remained the same when the AE was used while it

decreased from 5 to 3 when the DCNN was the base

trainer. The number of units at each layer showed a

variation. The DCNN’s number of layers remained

the same across RCs while the AE utilized 2 layers

with the same number of units at each layer for all

RCs.

Table 1: Best Parameters for RC and AE+RC models.

Table 2: Best Parameters for RC and DCNN+RC models.

4 CONCLUSIONS

In this research, we introduced our proposed 3D

spatiotemporal multi-view cross-learning framework

for the identification of individuals using EEG based

neural responses. We explored the effectiveness of

cross-model machine learning approaches compared

to regular classifiers. In addition, we investigated

individuals’ identifiability using the proposed

framework. The results indicate that the proposed

approach is promising, although more detailed

exploration is needed to achieve stable learning.

As an extension of this research, attention

mechanisms could be employed to enhance stability.

Furthermore, a longitudinal study involving data

collection over an extended period would help us

better understand the stability of EEG neural

responses.

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

946

Figure 6: Comparison of RCs when trained by base

learners, a) Regular classifiers, b) AE, c) DCNN (ACC

scores).

ACKNOWLEDGEMENTS

Research reported in this publication was supported

by an Institutional Development Award (IDeA) from

the National Institute of General Medical Sciences of

the National Institutes of Health under grant number

P20GM103424-20 via Louisiana Biomedical

Research Network.

Author Contribution Statement:

All authors discussed the methods, results, and

commented on the manuscript. Major individual

contributions are as in the following.

Figure 7: Effectiveness of cross-model learners across RCs

(ACC scores).

Ömer M. Soysal: Supervising the project,

conceptualization, pipeline design, supervising the

data collection, implementing the pipeline, leading

the preparation of the manuscript.

Iphy E. Kelvin: Pipeline implementation,

debugging, testing the code, data structure design,

running the code, assisting in the data collection and

writing the method section.

Esad M. Oztemel: Implementation of the

autoencoder and signal conditioning functions,

assisting in conceptualization, writing the

introduction, and results section.

Effectiveness of Cross-Model Learning Through View-Model Ensemble on Detection of Spatiotemporal EEG Patterns

947

Figure 8: Identifiability of individuals (AUC scores).

REFERENCES

Abdelhameed, A. M., & Bayoumi, M. (2018, December).

Semi-supervised deep learning system for epileptic

seizures onset prediction. In 2018 17th IEEE

international conference on machine learning and

applications (ICMLA) (pp. 1186-1191). IEEE.

Acharjee, R., & Ahamed, S. R. (2024, February).

Automatic Eyeblink Artifact Removal from Single

Channel EEG Signals Using One-Dimensional

Convolutional Denoising AE. In 2024 International

Conference on Computer, Electrical & Communication

Engineering (ICCECE) (pp. 1-7). IEEE.

Aggarwal, S., & Chugh, N. (2022). Review of machine

learning techniques for EEG based brain computer

interface. Archives of Computational Methods in

Engineering, 29(5), 3001-3020.

Ari, B., Siddique, K., Alçin, Ö. F., Aslan, M., Şengür, A.,

& Mehmood, R. M. (2022). Wavelet ELM-AE based

data augmentation and deep learning for efficient

emotion recognition using EEG recordings. IEEE

Access, 10, 72171-72181.

Arnau-González, P., Katsigiannis, S., Arevalillo-Herráez,

M., & Ramzan, N. (2021). BED: A new data set for

EEG-based biometrics. IEEE Internet of Things

Journal, 8(15), 12219-12230.

Atrey, P. K., Hossain, M. A., El Saddik, A., & Kankanhalli,

M. S. (2010). Multimodal fusion for multimedia

analysis: a survey. Multimedia systems, 16, 345-379.

Bandana Das, B., Kumar Ram, S., Sathya Babu, K.,

Mohapatra, R. K., & Mohanty, S. P. (2024). Person

identification using AE-CNN approach with multitask-

based EEG biometric. Multimedia Tools and

Applications, 1-21.

Bidgoly, A. J., Bidgoly, H. J., & Arezoumand, Z. (2020). A

survey on methods and challenges in EEG based

authentication. Computers & Security, 93, 101788.

Bharadwaj, S., Vatsa, M., & Singh, R. (2014). Biometric

quality: a review of fingerprint, iris, and face. EURASIP

journal on Image and Video Processing, 2014, 1-28.

Cimtay, Y., & Ekmekcioglu, E. (2020). Investigating the

use of pretrained convolutional neural network on

cross-subject and cross-dataset EEG emotion

recognition. Sensors, 20(7), 2034.

Dutt, M., Redhu, S., Goodwin, M., & Omlin, C. W. (2022,

October). Sleep Stage Identification based on Single-

Channel EEG Signals using 1-D Convolutional AEs.

In 2022 IEEE International Conference on E-health

Networking, Application & Services (HealthCom) (pp.

94-99). IEEE.

Emanet, F. Y., & Sekeroglu, K. Decoding Visual Stimuli

and Visual Imagery Information from EEG Signals

Utilizing Multi-Perspective 3D-CNN Based

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

948

Hierarchical Deep-Fusion Learning Network.

Proceedings Copyright, 381, 388.

Fidas, C. A., & Lyras, D. (2023). A review of EEG-based

user authentication: trends and future research

directions. IEEE Access, 11, 22917-22934.

Gao, Y., Fu, X., Ouyang, T., & Wang, Y. (2022). EEG-

GCN: spatio-temporal and self-adaptive graph

convolutional networks for single and multi-view EEG-

based emotion recognition. IEEE Signal Processing

Letters, 29, 1574-1578.

Gopal, S. R. K., & Shukla, D. (2021, August). Concealable

biometric-based continuous user authentication system

an EEG induced deep learning model. In 2021 IEEE

International Joint Conference on Biometrics

(IJCB) (pp. 1-8). IEEE.

Jia, Z., Lin, Y., Wang, J., Ning, X., He, Y., Zhou, R., ... &

Li-wei, H. L. (2021). Multi-view spatial-temporal

graph convolutional networks with domain

generalization for sleep stage classification. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 29, 1977-1986.

Khosla, A., Khandnor, P., & Chand, T. (2020). A

comparative analysis of signal processing and

classification methods for different applications based

on EEG signals. Biocybernetics and Biomedical

Engineering, 40(2), 649-690.

Kuncheva, L. I. (2014). Combining pattern classifiers:

methods and algorithms. John Wiley & Sons.

Mane, R., Robinson, N., Vinod, A. P., Lee, S. W., & Guan,

C. (2020, July). A multi-view CNN with novel variance

layer for motor imagery brain computer interface. In

2020 42nd annual international conference of the IEEE

engineering in medicine & biology society (EMBC)

(pp. 2950-2953). IEEE.

Mueller, S., Wang, D., Fox, M. D., Yeo, B. T., Sepulcre, J.,

Sabuncu, M. R., ... & Liu, H. (2013). Individual

variability in functional connectivity architecture of the

human brain. Neuron, 77(3), 586-595.

Nakamura, T., Goverdovsky, V., & Mandic, D. P. (2017).

In-ear EEG biometrics for feasible and readily

collectable real-world person authentication. IEEE

Transactions on Information Forensics and

Security, 13(3), 648-661.

Oh, S. L., Hagiwara, Y., Raghavendra, U., Yuvaraj, R.,

Arunkumar, N., Murugappan, M., & Acharya, U. R.

(2020). A deep learning approach for Parkinson’s

disease diagnosis from EEG signals. Neural Computing

and Applications, 32, 10927-10933.

Oztemel, M. E., & Soysal, Ö. M. (2024, April). Effect of

Signal Conditioning and Evoked-Potential Based

Representation on Stability and Distinctiveness of EEG

Brain Signatures. In 2024 12th International

Symposium on Digital Forensics and Security

(ISDFS) (pp. 1-7). IEEE.

Qi, H., Xue, Y., Xu, L., Cao, Y., & Jiao, X. (2018). A

speedy calibration method using Riemannian geometry

measurement and other-subject samples on a P300

speller. IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 26(3), 602-608.

Plucińska, R.; Jędrzejewski, K.; Malinowska, U.; Rogala, J.

Leveraging Multiple Distinct EEG Training Sessions

for Improvement of Spectral-Based Biometric

Verification Results. Sensors 2023, 23, 2057.

Riera, A., Soria-Frisch, A., Caparrini, M., Grau, C., &

Ruffini, G. (2007). Unobtrusive biometric system based

on electroencephalogram analysis. EURASIP Journal

on Advances in Signal Processing, 2008, 1-8.

Spyrou, L., Kouchaki, S., & Sanei, S. (2015, September).

Multiview classification of brain data through tensor

factorisation. In 2015 IEEE 25th international

workshop on Machine Learning for Signal Processing

(MLSP) (pp. 1-6). IEEE.

Tian, C., Ma, Y., Cammon, J., Fang, F., Zhang, Y., & Meng,

M. (2023). Dual-encoder VAE-GAN with

spatiotemporal features for emotional EEG data

augmentation. IEEE Transactions on Neural Systems

and Rehabilitation Engineering, 31, 2018-2027.

Waytowich, N. R., Lawhern, V. J., Bohannon, A. W., Ball,

K. R., & Lance, B. J. (2016). Spectral transfer learning

using information geometry for a user-independent

brain-computer interface. Frontiers in neuroscience, 10,

430.

Weng, W., Gu, Y., Guo, S., Ma, Y., Yang, Z., Liu, Y., &

Chen, Y. (2024). Self-supervised Learning for

Electroencephalogram: A Systematic Survey. arXiv

preprint arXiv:2401.05446.

Wu, D., Lawhern, V. J., Hairston, W. D., & Lance, B. J.

(2016). Switching EEG headsets made easy: Reducing

offline calibration effort using active weighted

adaptation regularization. IEEE Transactions on

Neural Systems and Rehabilitation

Engineering, 24(11), 1125-1137.

Wu, D., Xu, Y., & Lu, B. L. (2020). Transfer learning for

EEG-based brain–computer interfaces: A review of

progress made since 2016. IEEE Transactions on

Cognitive and Developmental Systems, 14(1), 4-19.

Xu, C., Tao, D., & Xu, C. (2013). A survey on multi-view

learning. arXiv preprint arXiv:1304.5634.

Yao, J., & Motani, M. (2018). Deep spatio-temporal feature

learning using AEs.

Zhao, M., Gao, H., Wang, W., & Qu, J. (2020). Research

on human-computer interaction intention recognition

based on EEG and eye movement. IEEE Access, 8,

145824-145832.

Zhou, T., & Wang, S. (2024). Spatio-Temporal Adaptive

Diffusion Models for EEG Super-Resolution in

Epilepsy Diagnosis. arXiv preprint arXiv:2407.03089.

Effectiveness of Cross-Model Learning Through View-Model Ensemble on Detection of Spatiotemporal EEG Patterns

949