A novel ensemble model for identification and classification of cyber harassment on social media platform

Schools and universities shuttered as a result of the worldwide COVID-19 pandemic lockdown, and student screen time skyrocketed. Since the programs are delivered online, a spike in social media use during lockdown resulted in many pupils becoming victims of cyberbullying, which includes criticizing one another, posting sexual comments on images of young ladies, and using fake accounts to bully others. Machine Learning (ML) and Natural Language Processing (NLP) techniques are being used in a growing body of work on automated cyberbullying detection. Different machine learning methods, however, are unable to converge to the requisite accuracy. Thus, numerous classifier systems known as “ensemble learning” are proposed in order to improve predictive performance by aggregating the predictions from various models. In our proposed system, we use a novel method of detecting online harassment (cyberbullying) on the Instagram dataset. The attributes of abusive words are initially analyzed from feature selection and pre-trained word embedding language models like Bidirectional Encoder Representations from Transformers (BERT) and Embeddings from Language Models (ELMO). A knowledge-based frequent pattern method is used to find the intention of the harasser and is created by the Knowledge-BERT (K-BERT). The unsupervised approaches such as Latent Semantic Analysis (LSA), Frequent pattern growth (FP-Growth), and a clustering technique K-Means. The results from the detection models are ensembled using Extreme Gradient Boosting (XGBoost) for classifying the categories of online harassment. The performance of the ensemble model is then cross-validated using machine learning metrics and compared with various existing techniques. An ensemble model performs better with a higher F1 score of 92.04% with less error rate in the classification of harassment categories.