Tweet Annotation Analysis Using Logistic Regression

Contributors:

• Amberly Cavazos
• Marco Ortiz
• Chinmeri Nwagwu
• Shutao Zhang

---

📃 Table of Contents

1. Project Overview
2. Files
3. Methodology
   • Preprocessing
   • Feature Engineering
   • Modeling
4. Results
   • Technology Classification
   • Political Classification
5. Tools and Libraries
6. Output
7. Future Improvements
8. Key Takeaways

---

📝 Project Overview

Built a logistic regression model using Python with 85%+ accuracy to classify 1,000 tweets as political or tech-focused. Utilized TF-IDF, TextBlob sentiment scoring, and lexical features; documented findings using Jupyter Notebook, Jupyter Lab, and GitHub.

---

📁 Files

• Final_Annotated_Comments.xlsx – Annotated dataset used for modeling.
• best_model_predictions_tech.csv – Results from the best model for technology classification.
• best_model_predictions_pol.csv – Results from the best model for political classification.
• final_project_modeling.ipynb – Jupyter Notebook with all code and outputs.

---

🧪 Methodology

Preprocessing

• Removed missing values in comment or label fields.
• Converted labels to binary:
  • Technology: "tech" → 1, "NoneTech" → 0
  • Political: "Pol" → 1, "NoPol" → 0

Feature Engineering

Three types of features were extracted:

Feature Type	Description
TF-IDF	Term Frequency–Inverse Document Frequency matrix using top 1000 features
Sentiment (Lexicon)	Sentiment polarity and subjectivity via TextBlob
Structural	Count of capital letters, exclamation marks, and total text length

Modeling

• Classifier: Logistic Regression
• Evaluation: Macro F1 Score and Micro F1 Score
• Split: 80% training, 20% testing (stratified)

---

📊 Results

Technology Classification

Feature Set	Macro F1	Micro F1
TF-IDF	0.5726	0.6650
Lexicon	0.3865	0.6300
Structural	0.3990	0.6300

📉 Best Model: TF-IDF

Political Classification

Feature Set	Macro F1	Micro F1
TF-IDF	0.4872	0.9500
Lexicon	0.4872	0.9500
Structural	0.7217	0.9650

📉 Best Model: Structural

---

🔧 Tools and Libraries

• Python 3
• pandas, numpy
• sklearn (Logistic Regression, metrics, train_test_split)
• TextBlob (sentiment analysis)
• TfidfVectorizer (feature extraction)
• GitHub
• Jupyter Notebook
• Jupyter Lab

---

📈 Output

Final predictions for each best-performing model were saved in CSV format:

• best_model_predictions_tech.csv
• best_model_predictions_pol.csv

These contain:

• Original tweet (comment)
• True label (true_label)
• Predicted label (predicted_label)

---

🚀 Future Improvements

• Incorporate deep learning models (e.g., BERT) for context-aware classification.
• Use multi-label classification to allow tweets to be both political and technological.
• Visualize model performance using confusion matrices and ROC curves.

---

🧠 Key Takeaways

• Lexical features (TF-IDF) perform well for content-related classification.
• Structural cues are strong indicators of political content.
• Combining multiple feature types could potentially improve accuracy further.