Early Behavioral Change Prediction on Social Media Through an Optimized Hybrid Deep Learning Ensemble Model
DOI:
https://doi.org/10.29284/75y3qz47Keywords:
Behaviour Change Prediction; Hybrid Deep Learning; Stacked Classification; CNN-BiLSTM-Attention; Grey Wolf Optimizer; Social Media Analytics.Abstract
The unchecked growth of social media usage has made platforms a rich but noisy source of behavioural signals that, if decoded early, can help identify psychological distress, radicalisation tendencies, and other harmful behaviour shifts before they escalate. This paper proposes a Stacked Classification Approach built on an Optimized Hybrid Deep Learning Model for the early prediction of behaviour changes from social media activity. The framework first extracts contextual textual embeddings and behavioural meta-features from user posts, then fuses them through a Convolutional Neural Network-Bidirectional Long Short-Term Memory-Attention (CNN-BiLSTM-Attention) block that captures both local linguistic patterns and long-range temporal dependencies. An Improved Grey Wolf Optimizer (I-GWO) tunes the hybrid network's hyperparameters to avoid manual trial-and-error and to escape local optima. The deep feature representation is then passed through a stacked ensemble in which CNN-BiLSTM-Attention, Random Forest, and XGBoost act as base learners and a Logistic Regression meta-learner combines their outputs into a final prediction. Experiments on a curated dataset of 39,940 annotated social media posts show that the proposed model achieves 96.8% accuracy, 96.1% precision, 95.9% recall, and a 96.0% F1-score, outperforming standalone CNN, LSTM, CNN-LSTM, and CNN-BiLSTM-Attention baselines by a margin of 4.7 to 15.4 percentage points. The results confirm that combining optimized hybrid deep feature learning with stacked ensemble classification substantially improves the early and reliable detection of behavioural change on social media.
References
1. Alharbi, M., & Alghamdi, N. (2021). Lexicon-based detection of depressive language on Twitter using support vector machines. Journal of Social Computing and Applications, 9(2), 112–124.
2. Chen, Y., Wang, H., & Li, X. (2022). A convolutional neural network approach for emotion classification on microblogging platforms. Information Processing & Management, 59(4), 102–115.
3. Devi, R., & Ramesh, K. (2022). Stacked ensemble learning for imbalanced social media classification tasks. Expert Systems with Applications, 198, 116–129.
4. Fathi, A., & Nasr, M. (2022). An improved grey wolf optimizer for hyperparameter tuning of deep neural networks. Applied Soft Computing, 118, 108–121.
5. Hussain, S., & Ahmed, F. (2023). Cross-platform behaviour prediction using stacked deep feature ensembles. IEEE Access, 11, 45210–45223.
6. Kumar, A., & Singh, P. (2023). Early depression detection from social media using LSTM-based sequence modelling. Computers in Human Behavior Reports, 10, 100–113.
7. Patel, D., Mehta, S., & Rao, V. (2024). Particle swarm optimized CNN-LSTM model for suicidal ideation detection on social platforms. Neural Computing and Applications, 36(3), 2201–2216.
8. Zhao, L., & Liu, J. (2023). Attention-augmented CNN-BiLSTM for cyberbullying detection in social networks. Knowledge-Based Systems, 262, 110–124.
Downloads
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
This work is licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
