skelouse/tweet-sentiment-analysis

Table of Contents

• 1  Model Selection for Tweet Sentiment Analysis
  • 1.1  This Repository Contains
  • 1.2  Questions
  • 1.3  Using the OSEMN Process
  • 1.4  Results
    • 1.4.1  Gradient Boosting Classifier
    • 1.4.2  Stacking Classifier
  • 1.5  Recommendations
  • 1.6  Next Steps
  • 1.7  Repository Structure

Model Selection for Tweet Sentiment Analysis

This Repository Contains

• A Jupyter notebook student.ipynb showing our tweet sentiment analysis
• The dataset itself Obtained from brands-and-product-emotions
• A PowerPoint presentation of the data

Questions

• Which model will perform best?
• How well will a model perform?
• What can androids or iphones improve in their products to reduce negative feedback?

Using the OSEMN Process

• Obtain the data
• Scrub the data
• Explore the data
• Model the data
• Interpret the data
• Reference

Results

Gradient Boosting Classifier

From the model:
• Slightly overfitting the training data
• Odd that the True Label(1) doesn't add up to 100%
• The best model overall at 79% accuracy
• Successfully predicting 64% and 82% of Negatives and Positives respectively

Stacking Classifier

From the model:
• Overfitting the training data at 98%
• Predicting 56% and 91% of Negatives and Positives Respectively
• 86% accuracy on the Testing data
• Still struggling to predict the negative emotion, slightly better than flipping a coin in that aspect

Recommendations

• Which model will perform best?
• Gradient Boosting for balanced recall, and stacking for the highest overall accuracy

• How well will a model perform?
• Overall Testing accuracy
  • Gradient Boosting: 76%
  • Stacking Classifier: 86%

• What can androids or iphones improve in their products to reduce negative feedback?
• Android
  • Contests
  • Events
• Iphone
  • Interface
  • Ease of use

Next Steps

• Build an sklearn pipeline with and grid search with the tokenizer and vectorizer parameters along with the parameters for each of the classifiers.
• Build neural networks for the data, using Oscar for hyperparameter tuning
• More tuning on the above models

Repository Structure

|   data
\--- tweet.csv
|
|   images
\--- phone.jpg
|
|   presentation.pdf
|   README.md
|   student.ipynb
|   readme.ipynb
|       
\--- styles
|      custom.css

\--- md
|      student.md
|      .png files of all graphs