Extra Trees leads 13-model cyberbullying benchmark on English tweets

A comparative study of 13 classification systems found that Extra Trees, operating on TF-IDF representations of text, delivered the strongest overall performance on a balanced, fine-grained cyberbullying dataset. The classifier achieved 97.79% accuracy, 97.82% precision, 97.79% recall and a 97.80% F1-score across six classes: age, gender, religion, ethnicity, other cyberbullying and non-cyberbullying. Support Vector Machine ranked second among the traditional models, while bidirectional long short-term memory and gated recurrent unit networks performed best among the deep-learning systems. The class-level results showed why a single headline score can be misleading: ethnicity was classified almost perfectly by Extra Trees, whereas gender generated more confusion across model types and age produced sharply different results for tree-based and recurrent architectures.

The research, published in Frontiers in Artificial Intelligence, evaluated models using the Fine-Grained Balanced Cyberbullying Dataset from IEEE DataPort. It contains more than 47,000 English-language tweets distributed approximately evenly across the six categories, with about 8,000 examples in each class. Abuse targeting age, gender, religion or ethnicity is separated from general cyberbullying and neutral posts. This balance limits the chance that a classifier will look effective simply by favoring a common label and makes comparisons between categories more informative. The study did not introduce a new culturally aware feature representation or a bias-mitigation algorithm. Instead, it established a common evaluation framework for asking whether models remain effective when online abuse is classified according to the identity or demographic focus involved.

Nine traditional machine-learning classifiers were trained on numerical vectors produced by Term Frequency–Inverse Document Frequency, or TF-IDF. The method emphasizes words that occur often within an individual post but are relatively uncommon across the full collection, making distinctive terms and short phrases useful signals for classification. The traditional systems were Support Vector Machine, Random Forest, Decision Tree, K-Nearest Neighbors, AdaBoost, XGBoost, Extra Trees, Logistic Regression and Gradient Boosting. Four neural architectures were tested on the same train-test partitions: TextCNN, bidirectional LSTM, bidirectional GRU and a hybrid CNN-BiLSTM. Unlike the sparse TF-IDF models, these deep networks learned dense word representations from text sequences, allowing them to represent relationships among words rather than relying only on explicit lexical markers.

Extra Trees produced the best reported aggregate result across the six-class task. Its advantage was particularly visible for religion- and ethnicity-related abuse. The model obtained an F1-score of 97.39% for religion and 100% for ethnicity, with both precision and recall reaching 100% in the latter category. Age was also classified accurately, with 98.61% precision, 98.02% recall and an F1-score of 98.31%. The support vector machine achieved 93.30% accuracy and a 93.28% F1-score overall, while random forest reached 91.66% accuracy. At the lower end of the traditional comparison, K-Nearest Neighbors recorded 63.71% accuracy and a 58.22% F1-score. The authors associated that result with the difficulty of identifying meaningful neighborhoods in high-dimensional sparse TF-IDF space, where Euclidean distances may not correspond closely to semantic similarity.

BiLSTM and BiGRU were nearly tied among the deep-learning systems. BiLSTM reported 90.50% accuracy and a 70.47% F1-score, while BiGRU achieved 90.49% accuracy and a 70.43% F1-score. Their bidirectional recurrent structures process a tweet in forward and backward directions, allowing each model to incorporate context supplied by words on both sides of a term. The results support the capacity of these architectures to learn sequential relationships, although neither surpassed the leading TF-IDF ensembles on the full multiclass task. TextCNN reached 78.75% accuracy and CNN-BiLSTM 81.09%. Their recall was 81.27% and 79.36%, respectively, while precision remained below 68%; the authors interpreted this pattern as more false-positive predictions and an emphasis on local rather than longer-range contextual patterns.

Class-level measurements revealed more variation than the aggregate rankings. Extra Trees performed especially well where offensive posts contained explicit identity-related words, consistent with the discriminative information available to TF-IDF. The recurrent networks were stronger on some context-dependent classes, including gender and non-cyberbullying, but their age results were markedly lower. BiLSTM achieved 62.77% precision, 67.01% recall and a 63.97% F1-score for age; BiGRU recorded 62.80% precision, 67.64% recall and a 64.10% F1-score. Confusion matrices showed recurring confusion between gender and age, whose messages can share stereotypes, indirect insults and colloquial expressions. By contrast, religion- and ethnicity-related posts often contained explicit discriminatory vocabulary that created clearer boundaries for lexical classifiers.

The results also expose a trade-off between predictive performance, computational demands and transparency. The TF-IDF ensembles required less training time and memory than the iterative neural models, and their individual feature contributions could be inspected through model coefficients or feature-importance scores. Deep networks offered a complementary ability to learn distributed representations of context, which can matter when abuse is implicit, sarcastic or expressed through a sequence of related terms. The authors did not claim that contextual learning alone solved culturally sensitive classification: both paradigms continued to struggle with subtle abuse, rapidly changing slang and expressions requiring pragmatic interpretation. Their findings instead point toward future testing of systems that combine interpretable lexical features with contextual neural representations; such hybrid methods were not evaluated in the present experiment.

Several limitations restrict how far the findings can be generalized. The corpus consists entirely of English-language Twitter posts collected in a particular temporal and social setting, so it does not represent multilingual communication, code-switching or the linguistic practices of all online communities. Labels for offensive language also depend partly on subjective human judgments, particularly for sarcasm, implicit aggression and coded expressions. The original dataset creators did not publicly report the cultural, linguistic or religious backgrounds of annotators, preventing the researchers from directly assessing how those perspectives may have shaped the labels. Although the six classes were balanced, the authors noted that abuse targeting minority communities or using culturally specific expressions may still be underrepresented. The comparison was also limited to lightweight recurrent and convolutional networks and did not include transformer-based systems such as BERT or RoBERTa.

For testing, the researchers randomly divided the data into 80% training and 20% testing subsets, stratifying the split by all six categories and fixing the random seed at 42. Text normalization and TF-IDF vectorization were fitted only on the training partition before being applied to the test data, limiting leakage from held-out examples. Support vector machines, random forests and Extra Trees were tuned with five-fold stratified cross-validation, whereas the remaining traditional systems and neural architectures used specified or empirically selected configurations. Deep models used 128-dimensional embeddings, Adam optimization, early stopping and a maximum of 20 epochs. These choices support reproducibility within the reported experiment, but they also mean the framework did not involve exhaustive tuning of every model. The clearest lesson is methodological: for culturally aware moderation, overall accuracy should not stand alone. Class-specific results and confusion patterns are needed to show whether a detector handles different kinds of abuse consistently, especially when one demographic category is easier to recognize than another.

Subject of Research: Comparative evaluation of machine-learning and deep-learning models for culturally sensitive cyberbullying detection across six demographic and non-demographic classes.

Article Title: Culturally aware comparative evaluation of machine learning and deep learning models for cyberbullying detection

Article References: Md Mohiuddin, G., Sayeed, M. S., & Yeng, O. L. (2026). Culturally aware comparative evaluation of machine learning and deep learning models for cyberbullying detection. Frontiers in Artificial Intelligence, 9, Article 1871932. https://doi.org/10.3389/frai.2026.1871932

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1871932

Keywords: cyberbullying detection, machine learning, deep learning, TF-IDF, Extra Trees, text classification, cultural sensitivity, demographic categories, social media, natural language processing, Culturally, aware