Intelligent NLP pipeline for detecting harmful online text with machine learning and a production-style Streamlit dashboard.
The Cyberbullying Detection System is an NLP-based machine learning project built to identify and classify cyberbullying-related text from social media-style content.
It combines text preprocessing, TF-IDF feature engineering, Logistic Regression classification, and an interactive Streamlit interface to deliver real-time predictions, model insights, and evaluation analytics.
This project is directly relevant to:
- online safety and moderation
- social media monitoring
- harmful content detection workflows
- AI-assisted trust and safety systems
Cyberbullying is a serious digital safety issue that affects mental health, user well-being, and platform trust.
Manual moderation alone is difficult to scale because:
- online content is generated at high volume
- harmful language can appear in many forms
- abusive text often needs fast intervention
- moderation teams need consistent support tools
Early detection of toxic and harmful language can help platforms build safer online environments and support more responsible content moderation.
- Build an end-to-end NLP pipeline for cyberbullying text detection
- Clean and transform raw text into machine-readable features
- Train a reliable classification model for harmful content prediction
- Evaluate performance using standard classification metrics
- Deploy the model in a Streamlit dashboard with a professional user experience
- Provide analytics, interpretability, and system diagnostics for demonstration and review
- Real-time text detection for instant cyberbullying classification
- NLP preprocessing pipeline for cleaning and normalizing text
- TF-IDF vectorization for converting text into model-ready numerical features
- Logistic Regression-based machine learning classification
- Premium multi-page Streamlit dashboard
- Analytics and visual insights for model performance and dataset behavior
- Model evaluation with accuracy, precision, recall, F1-score, and confusion matrix
- Explainability features showing influential terms behind predictions
- Python
- pandas
- numpy
- scikit-learn
- matplotlib
- seaborn
- plotly
- joblib
- nltk
- TF-IDF Vectorization
- Text preprocessing
- Lemmatization / stopword handling
- Streamlit
- Logistic Regression
project/
│── app.py
│── README.md
│── requirements.txt
│── runtime.txt
│
│── data/
│ │── aggression_parsed_dataset.csv
│ │── original_dataset_backup.csv
│
│── model/
│ │── cyberbullying_model_lr.pkl
│ │── tfidf_vectorizer.pkl
│ │── X_test_sparse.npz
│ │── y_test.npy
│
│── notebooks/
│ │── cyberbullying_detection_system.ipynb
│
│── scripts/
│ │── save_test_split.py
Note: In the current repository, several of these assets are stored at the project root for easier deployment with Streamlit.
- Used a labeled aggression / cyberbullying text dataset
- Prepared structured text-label data for supervised learning
- Removed URLs, mentions, punctuation, and noise
- Standardized text formatting for downstream processing
- Converted text to lowercase
- Removed stopwords
- Applied token normalization and optional lemmatization
- Transformed cleaned text into TF-IDF vectors
- Represented textual patterns as weighted numerical features
- Trained a Logistic Regression classifier
- Used a classical ML pipeline suitable for interpretable text classification
- Measured classification performance using:
- Accuracy
- Precision
- Recall
- F1 Score
- Confusion Matrix
- Built a Streamlit app for:
- real-time prediction
- analytics dashboard
- model diagnostics
- project presentation
- Accuracy: 93.6%
- Precision: 94.7%
- Recall: 92.3%
- F1 Score: 93.5%
- The model demonstrates strong overall performance for cyberbullying detection.
- High precision suggests the system is effective at reducing false positives.
- Strong recall indicates it can capture most harmful instances successfully.
- The balanced F1 score shows a robust trade-off between sensitivity and prediction quality.
The Streamlit application provides a polished multi-page dashboard for live prediction, analytics, system inspection, and project presentation.
Suggested screenshots to include:
- Overview page
- Live Detection page
- Analytics dashboard
- System Information page
- TF-IDF with Logistic Regression provides strong baseline performance for harmful text detection.
- Precision-focused performance is especially valuable in moderation support systems.
- Clean preprocessing has a major impact on text classification quality.
- A lightweight classical ML model can still deliver strong results when paired with thoughtful preprocessing and evaluation.
- Dashboard-based deployment significantly improves accessibility for non-technical users.
- Upgrade to BERT or Transformer-based architectures
- Improve dataset quality and class balance
- Add multi-class cyberbullying category prediction
- Introduce API integration for real-time moderation systems
- Expand explainability with SHAP or LIME
- Strengthen UI/UX for production-style deployment
- Add monitoring and drift analysis for model maintenance
git clone https://github.com/doyancha/Cyberbullying_Detection_System.git
cd Cyberbullying_Detection_Systempip install -r requirements.txtstreamlit run app.pyVisit the local Streamlit URL shown in the terminal, typically:
http://localhost:8501Mir Shahadut Hossain (doyancha)
Data Analyst | Machine Learning Enthusiast | NLP Project Builder
This project demonstrates:
- end-to-end NLP workflow design
- machine learning model development and evaluation
- practical deployment with Streamlit
- clean project presentation for a professional portfolio
It is designed to reflect a job-ready approach to data, machine learning, and product-facing analytics work.