StackOverflow Tag Prediction

This project predicts tags for StackOverflow questions using machine learning models, specifically designed to address the multi-label classification problem of assigning relevant tags based on the content of the question.

Dataset

• Source: Kaggle Facebook Recruiting III - Keyword Extraction competition.
• Features: Question titles, bodies, and associated tags.
• Size: 6 million rows of training data (questions with tags).

Problem Statement

• Objective: Predict tags for StackOverflow questions based on the question’s title and body.
• Challenge: Multi-label classification where each question can have multiple tags.

Techniques Used

1. Data Preprocessing:

   • Removed HTML tags and special characters.
   • Applied tokenization and stemming.
   • Vectorized text data using TF-IDF and CountVectorizer.

2. Clustering and Dimensionality Reduction:

   • Used Truncated SVD for dimensionality reduction on high-dimensional text data.

3. Models Implemented:

   • Logistic Regression with OneVsRestClassifier for multi-label classification.
   • SGDClassifier with L1 penalty for improved tag prediction accuracy.
   • Multilabel K-Nearest Neighbors (MLkNN) for tag prediction.

Results

• Evaluation Metrics: F1 score (micro and macro), Hamming Loss.
• Best Model Performance:
  • Macro F1 score: 0.77
  • Micro F1 score: 0.85
  • Hamming Loss: 0.14

Conclusion

The project effectively predicts relevant tags for StackOverflow questions by leveraging multi-label classification techniques, with optimized performance through data preprocessing and model selection.