Sub-event Detection in Twitter Streams - NLP Kaggle Competitition

---

CSC_51054_EP Machine and Deep Learning, Fall 2024

Team LLY Members: Ziyi LIU, Ling LIU, Yixing YANG

---

This repository contains the code for the Kaggle competition: Sub-event Detection in Twitter streams, a binary classification task.

Our team, "LLY", achieved 1st place on the private leaderboard and 12th place on the public leaderboard.

Project Goal

The goal of this project is to detect sub-events within Twitter streams related to specific main events. This involves processing tweets to identify and classify event-related information.

File Structure

.
├── data/                     # Contains raw and processed datasets (ignored by .gitignore)
│   └── challenge_data/
├── src/
│   ├── data.py               # Advanced data preprocessing and embedding generation
│   └── model.py              # RNN model implementation
├── challenge_data.py         # Initial script for data loading and basic preprocessing
├── cnn.ipynb                 # Jupyter notebook for CNN model experimentation
├── lstm.ipynb                # Jupyter notebook for LSTM model experimentation
├── rnn.ipynb                 # Jupyter notebook for RNN model experimentation
├── machine_learnings.ipynb   # Jupyter notebook for various traditional ML models
├── Logistic.ipynb            # Jupyter notebook for Logistic Regression model
├── voting.ipynb              # Jupyter notebook for combining models using a voting classifier
├── requirements.txt          # Python dependencies
├── DataChallengeReport.pdf   # Project report
├── INF554-Challenge-2024.pdf # Competition details
└── README.md                 # This file

Core Logic

Data Preprocessing

The primary data preprocessing is handled by src/data.py, with an initial version in challenge_data.py. Key steps include:

• Text Cleaning: Lowercasing, removing URLs, user mentions (while keeping hashtags), punctuation, and numbers.
• Language Handling: Translation of non-English tweets to English and emoji conversion to text.
• Normalization: Unicode normalisation and contraction expansion.
• Tokenization, Stopword Removal, and Lemmatization: Standard NLP techniques to prepare text for modelling.
• Parallel Processing: Utilized for efficient preprocessing of large datasets.
• Embedding Generation: src/data.py uses pre-trained GloVe embeddings (glove-twitter-200) to convert tweets into numerical vectors.

Models

Various models were explored and implemented:

• RNN: A custom RNNBinaryClassifier is defined in src/model.py using PyTorch.
• CNN, LSTM: Explored in their respective Jupyter notebooks (cnn.ipynb, lstm.ipynb).
• Traditional Machine Learning Models: Experiments with models like Logistic Regression are found in machine_learnings.ipynb and Logistic.ipynb.
• Voting Classifier: A voting.ipynb notebook details the combination of different models to improve performance.

Setup and Usage

1. Clone the repository:

   git clone https://github.com/llada60/Sub-event_Detection_in_Twitter_streams.git
   cd Sub-event_Detection_in_Twitter_streams

2. Install dependencies:

   pip install -r requirements.txt

3. Download Data: The data/challenge_data/ directory is excluded by .gitignore. You will need to download the competition data from Kaggle and place it in this directory.
4. Run Preprocessing: The src/data.py script can be run to perform the full preprocessing pipeline. It saves intermediate and final processed files.
5. Explore Notebooks: The Jupyter notebooks (*.ipynb) contain the model training, experimentation, and evaluation logic. Open and run these using Jupyter Lab or Jupyter Notebook.

Key Libraries

• PyTorch
• Gensim
• NLTK
• Pandas
• NumPy
• Scikit-learn
• Tqdm
• Emoji
• Langdetect
• Googletrans

---