Multi-Ideology Multi-Class Extremism Classification Using Deep Learning Techniques

Social media is an integral part of today’s social communication. Social media platforms have a global reach with immense popularity among the young generation. This reach and influencing power of social media has attracted extremist and terrorist organizations to social media platforms. Numerous terrorist organizations like ISIS, Taliban, Al-Qaeda, and Proud Boys and conspiracy theory groups like Alt-Right and QAnon spread their propaganda, radicalize and recruit youths via social media platforms. Thus, online extremism research is imperative to monitor extremists’ influence and their spread of hate on social media. The existing research is limited to the specific ideology, which results in a bias towards a particular ideology. The classification of extremism is presented only in binary or tertiary classes with no further insights. This research work presents the development of a seed dataset and balanced multi-ideology extremism text dataset with multi-class labels. Recently natural language processing with deep learning has gained significant attention in extremism detection research. Thus this research focuses on collecting, cleaning and classifying the extremist tweets. This study presents a multi-class classification of the balanced multi-ideology dataset. This dataset is termed Merged Islamic State of Iraq and Syria (ISIS) /Jihadist-White Supremacist (MIWS). The MIWS dataset is evaluated using pre-trained Bidirectional Encoder Representation for Transformers (BERT) and variants like Robustly Optimized BERT Pretraining Approach (RoBERTa) and DistilBERT and achieves the highest f1-score of 0.72. RoBERTa and DistilBERT provide f1-score of 0.68 and 0.71, respectively. Thus, deep learning can be effectively used to identify extremism from multiple ideologies and segregate them into propaganda, radicalization and recruitment.


I. INTRODUCTION
Social media has become an unstoppable force in recent years. Facebook, Twitter, and WhatsApp are the market leaders in social networks with a high userbase [1], [2]. Twitter has 206 million daily users, most in the 25-34 age group [2]. Facebook-owned services: Facebook, Messenger, The associate editor coordinating the review of this manuscript and approving it for publication was Patrizia Grifoni .
WhatsApp, and Instagram have 2.6 billion daily users [1]. Thus, these social media platforms have extensive reach among various nationalities, races, and ages. This has given different anti-social elements to spread their propaganda, radicalize and recruit people. Terrorist organizations such as ISIS, Taliban, and Al-Qaeda employ social media to attract vulnerable youth [3], [4], [5]. Similarly, far-right organizations like Proud Boys spread their propaganda via social media [6] Christchurch Mosque Tragedy [7] further cemented VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ the role of social media in extremism propagation where attackers live-streamed the attack. In recent, the capture of Kabul by the Taliban garnered much support on Twitter [8].
Online radicalization and hate speech influenced the Capitol Riots [9] and Poway Synagogue Shootings [10] In the recent time Buffalo Massacre 2022 [11], [12], the perpetrator believed in conspiracy theories like 'Great Replacement' spread on forums like 4chan [13] In addition, the Buffalo shooter was reported to be influenced by Christchurch and Oslo attackers [14]. The effect of online extremism can be felt in regions where extremist organizations don't even exist [15] Lone wolf attacks are the primary example of online radicalization [16] ISIS/Jihadist and White Supremacist ideologies are spread through social media platforms using the same strategies [17] Anti-government propaganda [18], hate and perpetuating violence against targeted people [19], community and nations are recurring themes in social media posts from ISIS/Jihadist and White Supremacist ideologies. Extremists also use antivaccine and COVID news to radicalize people [20] This proves extremists adapt their strategies to spread propaganda, radicalization, and recruit people based on current events. This makes online extremism research challenging. Online extremism research is the only way to analyze, predict, and restrict extremist ideologies on social media. Thus, it is necessary to understand propaganda, radicalization, and recruitment texts to counter the rise of extremism on social media platforms.
Propaganda is 'content, which can be biased and exploited for personal or political gains' [21] Misinformation is also a part of propaganda. Radicalization is a 'change in behaviour, attitude or perception towards person or community [22]'. Extremists radicalize people by misquoting religious texts, citing political uprisings, etc. Recruitment in the context of extremism is 'inciting people to join the terrorist cause or commit a violent attack' [23]. Organizations like ISIS, the Taliban, and Al-Qaeda recruit by glorifying the death of terrorists. At the same time, right-wing white supremacists use 'anti-government themes,' 'anti-Semitism,' and recently 'coronavirus themes' for recruitment [24].
As social media use increases exponentially, extremist organizations have spread their reach to every corner of the globe. Thus, it is imperative to develop automatic detection of propaganda, radicalization, and recruitment texts to stop the spread of online extremism.
The motivation behind this study is to propose an extremist identification system free from ideological bias. Thus any extremist content should be identified without any judgement on political and religious leanings. Another motivation is to devise a system which can monitor social media so violent events like US Capitol Riots, Christchurch Attack or France Teacher attack [25].
Machine Learning has been used for extremism detection and analysis since the 2010s. Multiple researchers efficiently used Machine Learning (ML) with different features for extremism detection on various social media and websites. This is discussed in Section 2. But these ML algorithms and features face critical problems like data limitations and context identification issues.
Researchers used Deep Learning techniques in every possible data science field. Deep learning techniques that use transformer modules can recognize context better than other algorithms. So, to better identify the context related to propaganda, radicalization, and recruitment, Bidirectional Encoder Representations for Transformers (BERT), RoBERTa, and DistilBERT are used in this study.
Following are the contributions of this study: • Construction of balanced multi-ideologies, multi-class, extremism seed dataset gathered from journal papers, newspapers, and websites.
• Development of balanced multi-ideology extremism dataset, which consists of recent tweets.
• Development of a framework using natural language processing (NLP) techniques for online extremism dataset creation, labeling, and validation of online extremism tweets and classifying them into Propaganda, Radicalization, and Recruitment.
• Evaluation of balanced multi-ideology extremism dataset using pre-trained neural networks such as BERT and its variants such as RoBERTa and DistilBERT.
• Comparative analysis of BERT with variants of BERT such as RoBERTa and DistilBERT. This study is divided into different sections. Section II discusses the literature analysed for this work. The methodology is described in Section III. Section IV, Empirical Analysis, provides experimentation details for this study. Section V deals with Results and Discussion about the experiments performed. Sections VI, VII and VIII discuss Limitations, Conclusion and Future Work, respectively.

II. LITERATURE REVIEW
Extremism on social media has been discussed in literature since the early 2000s [26] Studies from 2015 to 2021 are considered to keep the research relevant and up to date. Most studies discuss extremism for a single ideology [5], [27], [28], [29]. There are different extremist organizations with numerous ideologies. This makes the classification of extremism biased to a particular ideology. So, there is a need for analysis and classification research that incorporates multiple extremist ideologies.
There are only a handful of standard datasets. These datasets [30] and [31] deal with ISIS Tweets propagated during its heydays from 2014-2017. ISIS standard dataset contains nearly 15,000 extremist tweets. White Supremacist ideology standard datasets [28], [32] are from Stormfront and Gab websites, respectively. The problem with standard datasets is that they are outdated and unbalanced. Therefore, most extremist researchers prefer to gather their data. Most extremism research is carried out on data collected from Twitter [27], [33], [34] This is because of the popularity of Twitter among young people and its microblogging format. Few researchers have also explored Facebook [35], Stormfront [28], and Gab [32] for extremism detection. Twitter policies restrict researchers from publishing tweet data. In addition, extremist accounts may get suspended. Thus, the collection of extremism data is challenging.
Studies like [46] and [47] provide comparable results on SVM with a precision of 0.92 and an accuracy of 0.95, respectively. Random Forest in a few studies [43], [48] gives better results with an f1-score of 0.84 and 0.93, respectively. Deep learning techniques [28], [39] provide better results than machine learning algorithms. A study using BERT with Word2Vec [33] gives the f1-score of 0.79 and a precision of 0.80. Similarly, a study using LSTM [39] provides a precision of 0.8596, and LSTM + CNN [41] also offers an accuracy of 0.9266. Thus, it is observed that deep learning techniques are better than machine learning algorithms in performance without using explicit feature extraction methods.
Few studies use Cohen's Kappa (McHugh, 2012) and Fleiss's Kappa (Fleiss, 1971) for validation. These methods are used to validate the labels manually. Most works [39], [53] use two experts to verify that Cohen's Kappa ranges from 0.10 to 0.76. Fleiss's Kappa in [28] gives kappa around 0.4 to 0.6, with three experts for validation. The issues plaguing manual validation are expert bias, a small sample for validation, and low metrics for validation.
Recently more extremism detection studies have been published. This study [54] is focused on radicalization detection using sentiment analysis. The authors focus on ISIS/ Jihadist ideology. The authors use SenticNet, and AffectiveSpace as features while Logistic Regression and LibSVM as classifiers. This study also classifies text into binary classes: extremism or non-extremism and hate or non-hate.
Religious extremism in the Kazakh language is identified by [55] The extremism text is collected from Vkontakte [56] The authors identified extremism text from keywords like kafir, kill, etc. Thus, limiting the study to ISIS/Jihadist ideology. The annotation is performed by observing extremists' keywords in the collected corpus. This annotation is binary. This study used Logistic Regression, SVM, and Random Forest for extremism text classification.
Reference [57] another study used CNN and LSTM for extremism classification. The authors use the previously collected Vkontakte dataset for classification in this study. The authors got an AUC of 0.99 for CNN and LSTM extremism text classification. Table 1 provides details of the literature studied for extremism detection. Table 1 compares studies based on the source from which data is collected (Data Source), algorithms or techniques used, classes or labels with size, and different performance metrics used. Thus, it can be observed from Table 1 that Twitter is the most studied platform for extremism research. The binary classification dominates the extremism detection research. The performance metrics primarily used are precision, f1-score, and ROC-AUC. Similarly, these conclusions can also be obtained from the systematic literature review [58], covering multiple online extremism detection studies and their current issues.

III. METHODOLOGY
This section presents the data collection, labeling, and validation process. Figure 1 shows the entire process flow for creating, labeling, and classifying the ideologically balanced MIWS dataset. The process can be divided into five steps: Collection of Data, Validation of Seed Data, Merging, and Classification. The following sections explain the steps mentioned above: Four hundred examples within the seed dataset are selected from varied sources like newspapers, journal articles, websites, blogs, and reports. There are two hundred posts of ISIS/Jihadists and two hundred for White Supremacists.

2) SEED DATA COLLECTION
Examples from research articles and websites that identify influential propagandists, violent radicals, and extremist recruiters are collected for the seed dataset.

a: SOURCES
The first ideology considers for seed data collection is ISIS/Jihadist. The reason for selecting ISIS/Jihadist ideology is their vast extent of propaganda, targeted recruitment, and multiple violent acts. White Supremacist ideology is also considered for seed data collection for similar reasons. The motive of the seed dataset is to identify text for propaganda, radicalization, and recruitment. Different sources like newspapers and journal articles and counterextremism websites are selected. Snowballing technique is also used to gather text examples of propaganda, radicalization, and recruitment from verified sources.

b: JOURNAL PAPERS AND REPORTS
The seed data in journal papers or research articles explicitly is small in numbers. Thus multiple reports are also considered. Details about journal papers and reports are provided in this study [58] The search for journal papers was limited

1) TWEET COLLECTION
A total of 60,000 tweets are included in the MIWS dataset, of which 30,000 belong to ISIS/Jihadist, and 30,000 belong to White Supremacist ideology.
2) DATA LABELLING NLP techniques like topic modelling and word distances are used to label the data. Latent Dirichlet Allocation (LDA) [67] is used for topic identification. Similarly, Word Mover's Distance (WMD) [68] is employed to determine the similarity between topics and classes.

3) LDA ON COLLECTED TWEETS
LDA provides topic-wise distribution of tweets. LDA determines these topics based on influencing words within the documents. Thus important words get highlighted within a specific topic. ISIS/Jihadist and White Supremacist tweets were divided into three LDA topics. Hyperparameter tuning was performed to produce the best model.

4) COMPARISON OF LABELED SEED DATASET AND LDA TOPICS OF COLLECTED TWEETS USING WORD MOVER'S DISTANCE
The LDA topics from collected tweets are compared with seed labels using Word Mover's Distance. This process is performed individually on each ideology. This is because words in ISIS /Jihadist and White Supremacist extremists significantly differ in the seed dataset. Word Mover's Distance (WMD) is applied to validate and label the topics from seed labels. WMD provides purposeful differentiation between terms from their co-occurrences within the corpus. So similarity increases when the distance between words decreases. WMD was applied using word2vec, which was pretrained on Google News Vector. Algorithm 1 describes the process in detail. Here seed dataset (Sd) contains three labels Propaganda (P), Radicalization (Rd), and Recruitment (Rc). Ct denotes collected tweets. Ct is passed to the LDA algorithm with the number of classes (n). LDA segregates tweets into T1, T2 and T3 topics. Each label text from Sd is compared with each tweet with the topic in Ct. This comparison is performed using WMD, which provides distances between texts. The computational cost of the algorithm depends upon the number of labels in Sd and the number of topics in Ct. Therefore, the algorithm has O(N 2 ) time complexity.
The average distance between label and topic was recorded, and the corresponding label whose average distance is least was assigned to that topic. Table 3 and Table 4 provide WMD distance between topics and seed labels; only the least average distances are considered as words should be near each other. Thus the lowest distance points to the label assigned to the topic. The radicalization label was given to Topic 0, which has the lowest WMD distance of 0.8575; Propaganda was assigned to Topic 1 as the lowest distance is 0.8455, and Recruitment was assigned to Topic 2 as the lowest distance of 0.8464 for ISIS/Jihadist collected tweets as seen in Table 3. Similarly, the Propaganda label was given to Topic 2 with the lowest distance of 0.7924, Radicalization to Topic 1 with the lowest distance is 0.8021, and Recruitment for Topic 0 with the lowest distance is 0.8032 for White Supremacist collected tweets, as seen in Table 4. This whole process is explained in this study [69].
The correctness labels were manually verified by taking random samples and checking the words belonging to Propaganda, Radicalization and Recruitment, as mentioned in [58].

5) MIWS DATASET
Topic 0 from Table 2 and Topic 1 from Table 4 are labeled as Radicalization. Radicalization has 18,120 tweets which are usually politically aligned. Topic 1 from Table 2 and Topic 2  from Table 3 are assigned Propaganda labels. Propaganda with 24,987 tweets is the largest of all three classes and is oriented towards religion, achievements, and glorification of ideology. Topic 2 from Table 3 and Topic 0 from Table 4   Propaganda, Radicalization and Recruitment was calculated. These TFIDF scores were used as paired input to t-test algorithm. Two hypotheses were generated: H0 -There no significant difference between classes H1 -There is significant difference between classes Table 5 provides details about classes and their p-value. As per hypotheses, if p < 0.05, H0 is reject, hence there is significant difference between classes. The p-value after t-test between Propaganda, Radicalization and Recruitment are less than 0.05. So H0 is rejected. Thus the difference between classes is statistically significant.
After this annotation, both ideologies, ISIS/Jihadist and White Supremacist are merged. This new dataset is called Merged ISIS/Jihadist-White Supremacist (MIWS). Table 5, provides descriptions of Seed and MIWS datasets. Table 5 shows the timespan of data collection, number of classes, features, word count, and size of classes.

7) BALANCING DATASETS
Propaganda tweets are more than Radicalization and Recruitment tweets. This affects the overall accuracy of the model. So, to balance the data, Randomized Under Sampling was used [70]

IV. EMPIRICAL ANALYSIS
This section provides information about deep learning classifiers, hyperparameters, and results for balanced MIWS extremism classification into propaganda, radicalization, and recruitment. VOLUME 10, 2022

A. EXPERIMENTAL SETUP
The experiments were conducted on two different systems. The initial experiments were conducted on HP Workstation Z8 G4 machine with Xeon 3GHz processor, 128 GB RAM, and Nvidia Quadro P400 2GB GPU. Hyperparameter tuning experiments were performed on Google Colab Pro.

B. DEEP LEARNING CLASSIFIER 1) BERT
Bidirectional Encoder Representations for Transformers (BERT) use the transformer attention model with the help of encoders [71] BERT processes text sequences in both left-right and right-left ways. This makes BERT efficient and accurate. BERT's stellar performance in NLP problems was the reason to choose BERT as a classifier for online extremism detection research. Figure 2 describes the general architecture of BERT used in this study. Table 6 shows hyperparameters used to obtain results using BERT. Different pre-trained networks were utilized along with BERT, such as Bert-base-uncased, Bert-large-uncased, Bert-base-cased, Bert-base-twitter, and Bert-large-twitter. The 'base' pretrained networks usually consist of 12 layers and 768 hidden units, while 'large' pre-trained networks have 24 layers and 1024 hidden units.

2) DISTILBERT
DistilBERT is another variant of BERT [72] The DistilBERT model was created using the knowledge distillation process. This makes DistilBERT smaller and faster than BERT. For this study, DistilBERT is used in combination with pretrained Bert-base-uncased. As seen in Figure 2, the architecture of DistilBERT is similar to BERT only difference between them is that DistilBERT performs knowledge distillation while pre-training. As seen in Table 7, fewer hyperparameters were used during DistilBERT training due to early convergence to an acceptable result.

3) ROBERTA
RoBERTa [73] enhances BERT's masking strategy and removes the next sentence prediction embedded in BERT. This can be observed in Figure 3. RoBERTa was modeled by training BERT on large mini-batches and different learning rates. RoBERTa was also trained on larger datasets for a longer time, making RoBERTa more generalized for downstream tasks. As seen in Table 8, RoBERTa was finetuned using the same hyperparameters as BERT. The pre-trained network used is Roberta-base-uncased, which has 12 layers and 768 hidden units.

V. RESULTS AND DISCUSSIONS
Earlier literature has used ROC-AUC, Precision, Recall, f1-score, and Accuracy as the performance metric. We have selected Precision, Recall, and f1-score as performance metrics. The reason behind choosing these performance metrics is to achieve higher performance using class balance and obtain the correct representation of misclassified elements. Seven combinations of Deep Learning Classifier and pre-trained network were performed to achieve significant results in Table 9. BERT and variants give acceptable results for MIWS classification. The precision for BERT and variants lies between 0.69 to 0.72, recall lies between 0.68 to 0.72, and the f1-score ranges from 0.68 to 0.72. The most important results can be seen when BERT with bert-basetwitter and BERT with bert-large-twitter are used. BERT with bert-base-twitter gives precision, recall, f1-score and accuracy of about 0.71, while BERT with bert-large-twitter gives precision, recall, f1-score and accuracy of about 0.72 as seen in Figure 4, Figure 5, and Figure 6. DistilBERT with bert-base-uncased offers similar results to BERT with bertbase-Twitter but doesn't hold up in label-wise comparison. BERT results are better than other combinations because pretrained networks are trained on Twitter datasets.
In Table 10, the rank significance of the results is provided. The ranks were calculated in descending order; thus, the lower the rank higher the significance. The Friedman Rank Test shows a p-value = 0.0006, which is less than 0.05. Therefore, the results obtained are significant for the collected dataset. The BERT model with Twitter pretrained has a lower rank compared to other models. Thus Table 11 and Table 12 show label-wise evaluation for BERT. It can be seen in both cases; performance is similar for propaganda and radicalization. But performance suffers for recruitment class for both experiments. This can be because recruitment may have more similar words to other classes.

VI. LIMITATIONS
Few limitations can be addressed in the future.
• The ideologically balanced MIWS dataset is labeled from the seed dataset with only 400 examples of ISIS and White supremacist ideology. In the future, the size of the seed dataset can be increased.
• Seed dataset contain event specific texts such as ISIS's threats to America after bombing etc. Thus only limited extremist keyword were present in seed dataset.
• The custom data collected was based on keywords. These keywords are prominently used by extremists but could have been used by critics, journalists, or news VOLUME 10, 2022   agencies to report the news. Thus, some false positives are there in the MIWS dataset.
• The tweet data collected is also event specific. Most jihadist tweets were about the Taliban takeover of   Afghanistan, while white supremacists are antisemitic and anti-vaccine. This may cause event-related classification issues in the trained model.
• Comparison using WMD gives close distances between classes; this can be due to similar keywords across categories. So, the misclassification of some custom-collected tweets may have happened.
• Some tweets collected were unavailable on Twitter due to their extremist nature. Thus tweet recollection or user account monitoring was hampered. VOLUME 10, 2022

VII. CONCLUSION
This research presents the development of a seed dataset and balanced multi-ideologies, multi-class, extremism tweet dataset and its evaluation using BERT and its variants. This work is the first to classify extremist tweets into propaganda, radicalization, and recruitment using deep learning to the best of our knowledge.
Examples of propaganda, radicalization and recruitment texts are collected from research articles, blogs, and websites. Thus, varied samples of seeds are collected to reduce researcher bias. Custom tweets are gathered using important extremism-related keywords from research articles, blogs, and websites. A total of 60,000 tweets from ISIS/Jihadist and White Supremacist ideologies were collected to create a balanced multi-ideology online extremism dataset. NLP techniques such as topic modelling, i.e., LDA and word distances, i.e., WMD are used for labelling collected tweets. These methods ensure tweets are segregated into propaganda, radicalization and recruitment.
This research work evaluates the multi-ideology, multiclass extremism MIWS dataset using BERT and its variants such as RoBERTa and DistilBERT. RoBERTa offers a precision of 0.69, recall of 0.68, and f1-score of 0.69. DistilBERT gives a precision of 0.71, recall of 0.71, and f1-score of 0.71.
In this research work, two different BERT pre-trained networks are employed. The first pre-trained network is trained on Wikipedia data, while the second is trained on Twitter data. Pre-trained network BERT trained on Twitter data provides the best results showing precision of 0.72, recall of 0.72, and an f1-score of 0.72.
This study sets objectives to detect extremism with insights. The automatic detection of extremism with insights can help to analyze extremist threats rapidly. This study can be used by researchers, social media networks, social media analysts, and government agencies to detect extremist propaganda and recruitment. Thus, social media networks or government agencies can take preventive measures to curb the spread of online extremism.

VIII. FUTURE WORK
This research has opened multiple avenues to classify better and detect online extremism. Following are some areas that can be enhanced in the future: • Increasing the size of the Seed Dataset: More seed examples identified as propaganda, radicalization, and recruitment can be included.
• Geographical Location as Feature: SpaCy can extract location from tweets. These geographical locations can be used as a feature to classify extremist tweets.
• Real-Time Classification: Tweets can be fetched in real-time and classified into propaganda, radicalization, and recruitment. He has expertise and experience in cutting-edge research and AI and deep learning projects for the last more than 25 years. He has published more than 100 widely in many excellent peer-reviewed journals on various topics ranging from cutting-edge AI, education policies, teaching-learning practices, and AI for all. He has also published three patents and delivered keynote speeches at various national and international forums, including at Machine Intelligence Laboratory, USA; IIT Bombay under the World Bank Project; the International Indian Science Festival organized by the Department of Science Technology, Government of India; and many more. He was a recipient of the two SPARC projects worth INR 166 lacs from MHRD, Government of India in AI, in collaboration with Arizona State University, USA; and the University of Queensland, Australia. He was also a recipient of numerous prestigious awards, such as Erasmus+ Faculty Mobility Grant to Poland, DUO-India Professors Fellowship for research in responsible AI in collaboration with Brunel University, U.K., LEAP Grant at Cambridge University, U.K., UKIERI Grant with Aston University, U.K., and the Grant from Royal Academy of Engineering, U.K., under Newton Bhabha Fund. He is also an Academic Editor of the PeerJ Computer Science journal and an Associate Editor of IEEE ACCESS journal.
AJITH ABRAHAM (Senior Member, IEEE) received the M.S. degree from Nanyang Technological University, Singapore, in 1998, and the Ph.D. degree in computer science from Monash University, Melbourne, Australia, in 2001.
He is currently the Director of the Machine Intelligence Research Laboratories (MIR Laboratories), a Not-for-Profit Scientific Network for Innovation and Research Excellence Connecting Industry and Academia. The Network with HQ in Seattle, USA, is currently has over 1,500 scientific members from over 105 countries. As an Investigator/a Co-Investigator, he has won research grants worth over 100 Million USA Dollar. He currently holds two university professorial appointments. He works as a Professor in artificial intelligence with Innopolis University, Russia, and the Yayasan Tun Ismail Mohamed Ali Professorial Chair in artificial intelligence with UCSI, Malaysia. He works in a multi-disciplinary environment. He has authored/coauthored more than 1,400 research publications out of which there are more than 100 books covering various aspects of computer science. One of his books was translated into Japanese and a few other articles were translated into Russian and Chinese. He has more than 46,000 academic citations (H-index of more than 102 as Per Google Scholar). He has given over 150 plenary lectures and conference tutorials (in more than 20 countries). He was the Chair of IEEE Systems Man and Cybernetics Society Technical Committee on Soft Computing (which has over 200 members), from 2008 to 2021, and served as a Distinguished Lecturer for IEEE Computer Society representing Europe, from 2011 to 2013. He was the Editor-in-Chief of Engineering Applications of Artificial Intelligence (EAAI), from 2016 to 2021, and serves/served on the editorial board for over 15 international journals indexed by Thomson IS. VOLUME 10, 2022