Disaster Tweets Classification

https://www.kaggle.com/code/kyutayasuda/disaster-tweet

Project Description

This project aims to classify tweets as related to disasters or not using machine learning. By leveraging a Random Forest model combined with text preprocessing techniques, the model predicts whether a tweet indicates a disaster-related event.

Key Highlights

1. Dataset: A labeled dataset containing tweets with information about their disaster relevance (target: 0 or 1).

2. Objective: Build a machine learning pipeline to preprocess tweet data and classify tweets as disaster-related or not.

3. Tools and Libraries:

   • Python

   • Libraries: pandas, scikit-learn, and TfidfVectorizer for text preprocessing, model training, and evaluation.

   • Jupyter Notebook for implementation.

Project Workflow

1. Data Preprocessing:

   • Utilized TfidfVectorizer to transform textual features (text, keyword, location) into numerical vectors.

   • Applied feature extraction with a maximum of 1000 features for text and 100 features for keyword and location.

   • Handled missing values and normalized text data.

2. Feature Engineering:

   • Created a preprocessing pipeline using ColumnTransformer for efficient feature handling.

3. Model Implementation:

   • Used a Random Forest Classifier with specific hyperparameters:

     • Random State: 42

   • Integrated the preprocessing pipeline with the classifier using scikit-learn's Pipeline.

4. Cross-Validation:

   • Performed K-fold cross-validation with k=5 to assess model stability and performance.

   • Results:

     • Mean Accuracy: 77.58%

     • Standard Deviation: 1.30%

5. Evaluation:

   • Evaluated the model on a test set, achieving the following metrics:

     • Accuracy: 78%

     • Precision (Class 0): 77%

     • Recall (Class 1): 66%

     • F1-Score (Class 1): 72%

Learning Outcomes

1. Developed a complete pipeline for text preprocessing and model training.

2. Enhanced understanding of text feature extraction techniques such as TF-IDF.

3. Learned to integrate cross-validation for robust model evaluation.

4. Practiced exporting and visualizing prediction results in a Kaggle-style competition.

Next Steps

• Experiment with advanced models like Gradient Boosting or deep learning (e.g., LSTM, BERT) for improved accuracy.

• Perform hyperparameter tuning using Grid Search or Bayesian Optimization.

• Include additional features such as word embeddings for better semantic understanding.