Multi-Label-Toxic-Comment-Classification-using-TF-IDF-and-Logistic-Regression

“Built a multi-label toxic comment classification system using TF-IDF features and a One-Vs-Rest Logistic Regression model, evaluated with F1-score and ROC-AUC to handle class imbalance.”

📌 Project Overview :

Online platforms often struggle to moderate toxic and abusive user-generated content at scale. This project addresses that challenge by building a multi-label toxic comment classification system using classical Natural Language Processing (NLP) and machine learning techniques.

The model is designed to identify multiple types of toxicity within a single comment, including toxic, severe toxic, obscene, threat, insult, and identity hate. The project emphasizes a clean end-to-end machine learning workflow with a strong focus on feature engineering, interpretability, and proper evaluation for imbalanced data.

---

🧠 Problem Statement :

Toxic comments can negatively impact online communities and user experience. Since a single comment can belong to multiple toxicity categories, this problem is framed as a multi-label text classification task rather than a simple binary classification problem.

Key challenges addressed:

• Multi-label classification
• Highly imbalanced target classes
• Converting unstructured text into meaningful numerical features
• Using appropriate evaluation metrics beyond accuracy

---

📂 Dataset Description :

The dataset consists of user comments along with binary labels indicating different types of toxicity:

• toxic
• severe_toxic
• obscene
• threat
• insult
• identity_hate

Each comment may belong to zero, one, or multiple labels, making the task multi-label in nature.

---

🔧 Methodology :

1️⃣ Data Preprocessing :

• Loaded and inspected the dataset
• Handled missing values in text fields
• Separated input features (comment_text) and target labels

2️⃣ Feature Engineering :

• Extracted linguistic features such as:

  • Comment length
  • Capital letter ratio
  • Average word length
  • Punctuation counts

• Converted raw text into numerical features using TF-IDF vectorization

  • Used unigrams and bigrams
  • Removed common English stop words

3️⃣ Model Development :

• Implemented a One-Vs-Rest Logistic Regression classifier to handle multi-label prediction
• Each toxicity label is modeled using a separate binary classifier
• This approach provides a strong, interpretable baseline for text classification tasks

---

📊 Model Evaluation :

Due to the imbalanced nature of toxic comment data, accuracy alone is not a reliable metric. The model is evaluated using:

• Precision, Recall, and F1-score (per label)
• Micro F1-score for overall performance
• Macro F1-score to treat all labels equally
• ROC-AUC (micro-average) using predicted probabilities

These metrics provide a more realistic understanding of model performance across all toxicity categories.

---

✅ Results & Observations :

• The model successfully learns patterns associated with different toxic behaviors
• Frequent toxicity categories (e.g., toxic, insult) perform better than rare ones (e.g., threat)
• TF-IDF with Logistic Regression provides a strong and interpretable baseline
• Proper metric selection significantly improves evaluation reliability

---

⚠️ Limitations :

• The model relies on surface-level text features and may struggle with:
  • Sarcasm
  • Context-dependent toxicity
  • Implicit or subtle hate speech
• No deep learning or contextual embeddings are used in this implementation

---

🔮 Future Improvements :

• Perform detailed error analysis on misclassified samples
• Tune decision thresholds for better recall/precision trade-offs
• Explore advanced NLP techniques such as word embeddings or transformer-based models
• Deploy the model as a REST API or web application for real-time moderation

---

🧰 Tools & Technologies Used :

• Python
• Pandas, NumPy
• Scikit-learn
• TF-IDF Vectorizer
• Logistic Regression (One-Vs-Rest)
• Jupyter Notebook

---

🎯 Key Takeaway :

This project demonstrates a complete and correct classical NLP machine learning pipeline for multi-label toxic comment classification, focusing on clarity, correctness, and industry-standard evaluation practices rather than unnecessary model complexity.

---

⚙️ Installation :

Clone the repository and install the required dependencies using:

pip install -r requirements.txt