StackOverflow Tag Predictor

Project Overview

This natural language processing project addresses the multi-label classification challenge on StackOverflow. The model analyzes the textual content and code blocks of a query to predict which tags (e.g., python, sql, multithreading) should be assigned to it, enabling automatic categorization and improved search navigation.

Key Features & Objectives

• Multi-Label Target Space: Predicts multiple tags concurrently for a single input question.
• Mixed Text Processing: Sanitizes and parses standard English sentences alongside embedded raw programming code blocks.
• SGD Linear Classifiers: Utilizes high-performance linear models to process high-dimensional word vectors with low overhead.

Technical Stack & Technologies

• Language: Python
• NLP & Vectorizers: Scikit-learn (TF-IDF vectorizer), NLTK, BeautifulSoup (HTML parser)
• Classifiers: SGDClassifier, OneVsRestClassifier
• Data Science: Pandas, NumPy

Core Techniques & Algorithms

• Multi-Label Vectorization: Constructed n-gram TF-IDF pipelines to extract text features, giving higher weights to tags and key programming keywords.
• Binary Relevance (One-vs-All): Leveraged OneVsRestClassifier wrappers to train separate binary classifiers for each tag label in the target dataset.
• Model Optimization & Jaccard Loss: Tuned Stochastic Gradient Descent (SGD) classifiers under modified huber and log losses, evaluating accuracy via micro F1-score and Jaccard similarity metrics.