NLP-Sentiment-Analysis: US-Airline Tweets

This repository contains a comparative study of various machine learning classifiers for natural language processing (NLP) sentiment analysis. The project evaluates model performance on a dataset of user opinions about US Airlines from X (formerly Twitter).

Project Overview

Objective: Classify the polarity of airline tweets using text classification techniques.

Dataset: Twitter_US_Airline_Sentiment.csv (attached to this repository)

Feature Extraction: TF-IDF vectorization tested under three vocabulary constraints (minimum 5 document frequency, max 2500 words, and max 500 words).

Models Evaluated:

• Logistic Regression
• Support Vector Machines (LinearSVC)
• Random Forests
• Feed-forward Neural Network

The classifiers are evaluated using 5-fold cross-validation. Performance is measured across three primary metrics:

• Accuracy
• F1-score
• Fit time

Results

• (Logistic Regression) consistently outperformed the others. It achieved its highest performance in Experiment 1, reaching 78.37% Accuracy and a 76.95% F1-score, while remaining highly computationally efficient.
• (Support Vector Machines (SVM)) delivered very competitive accuracy and F1-scores, and stood out by having the fastest fit times across all experiments.
• (Random Forest) and the (Feed-Forward Neural Network) yielded significantly lower accuracy and F1-scores, alongside much slower training times. The Neural Network, due to its complexity, was the least effective and least efficient model for this specific setup.

The Impact of Vocabulary Size (TF-IDF)

The experiments revealed a clear trade-off between computational cost and classification performance based on vocabulary constraints:

• Experiment 1 (min_df=5): Highest accuracy and F1-scores. Best choice for maximizing predictive performance.
• Experiment 2 (max_features=2500): Offers the best balance of solid performance and faster training times.
• Experiment 3 (max_features=500): Fastest to train, but suffers a significant drop in accuracy and F1-score due to the restricted vocabulary.

Note: See the attached Jupyter Notebook for full code, metrics, and exploratory data analysis (EDA).