BOT-InFUSED2: a transfer learning-based model for enhanced bot detection on X using instruction-fine-tuned stacked embeddings and a deep neural network

Accurate bot detection on microblogging platforms such as X (formerly Twitter) remains a significant and challenging problem, especially in the face of an expanding user base. Most existing methods are limited in their capacity to generalise well to new and unseen datasets due to the multifaceted nature of bot accounts, including their inherent adaptability and heterogeneity, which allow them to subtly evade detection. This paper proposes BOT-InFUSED2, a transfer learning-based ensemble model for enhanced bot detection. BOT-InFUSED2 combines vertically concatenated feature vectors from three pretrained embeddings (INSTRUCTOR, Universal Sentence Encoder, and DistilBERT) derived from tweets, together with account metadata and derived features. The resulting ensemble feature vector is then fed into a deep neural network for final classification. To avoid inconsistencies in detection, we average tweet-level bot scores so that each user receives a single bot score. To ensure high generalisability, BOT-InFUSED2 is validated beyond its training data through domain adaptation on unseen X data related to four seasons of the Big Brother Naija reality TV show. Experimental results show that BOT-InFUSED2 outperforms existing methods, achieving accuracy of 97.3%, F1-score of 0.972, and ROC-AUC of 0.973. These results highlight the potential of combining diverse pretrained embeddings and account features for accurate bot detection, paving the way for a more trustworthy X and an enhanced user experience. Future work will explore incorporating multimodal features and multilingual capabilities into BOT-InFUSED2 to achieve further accuracy gains and greater generalisability.