Table 2.

Popular machine learning techniques employed in online extremism detection.

Technique used for extremism detection	Study	Hyperparameter	Features	Performance metric	Remark
Naïve Bayes/multinomial Naïve Bayes	[35, 42, 46, 60–63]	Alpha = 0.01 [46]	n-grams, TF-IDF, and word2vec	Accuracy = 0.66 [46] and correctly classified instances = 89% [42]	Naïve Bayes or multinomial Naïve Bayes is used so as to build a model based on a probabilistic learning approach [46]
KNN	[64]	Distance = euclidean distance, K = 100 [64]	Term frequency	Precision = 0.48 and accuracy = 0.90 [64]	Distance-based approach for similarity in extremism text
Logistic regression	[36, 62, 65]	NA	Word2vec, fasttext, GloVe, and LIWC	F1 score = 99.77 [36] and accuracy = 0.70 [62]	Used for binary classification of extremism text
SVM	[36, 42, 44, 46, 49, 61–64, 66, 67]	Penalty = L1, tol = 1e − 3 [46]	n-grams, TF-IDF, word2vec, fasttext, GloVe, and PCA	Accuracy = 84 [67] and precision = 84 [49]	SVM segregates data using hyperplanes, so that classification is better. [46]
Random forest	[61, 63, 66, 68, 69]	Estimators = 100, Kfold = 5 [68], estimators = 100, max_depth = 50 [66]	n-grams, TF-IDF, word2vec, and GloVe	Accuracy = 100 [66] and F1-score = 0.93 [69]	Random forest is scalable and unaffected by outliers in extremism text dataset [66]
AdaBoost	[41, 42, 70]	Boosting iterations = 300 [41]	n-grams	Precision = 0.88 [70] and accuracy = 99.5 [42]	AdaBoost improves performance by combining weak classifiers
XGBoost	[59, 71]	Regularization = L2	Betweenness centrality and page rank	ROC-AUC curve = 0.95 [59]	XGBoost improves performance with faster learning