📧 Email Classification Using RNN & LSTM

Neural Networks & Fuzzy Logic (NNFL) Project

This project implements an end-to-end email classification system using Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks. The system classifies emails as ham (legitimate) or spam using deep learning.

The project includes:

• Dataset loading & cleaning
• Exploratory Data Analysis (EDA)
• Text preprocessing
• RNN and LSTM model training
• Evaluation (accuracy, confusion matrix, precision, recall, F1)
• Custom email prediction pipeline

🚀 Features

• Raw email text preprocessing (tokenization, stemming, stopword removal, n-grams)
• Train/validation/test split with stratification
• Deep learning-based text classification
• RNN baseline model
• LSTM enhanced model
• Model checkpointing + early stopping
• Confusion matrix & metrics visualization
• Custom email prediction function
• Ready for deployment/inference

🧠 Project Architecture

Email_Classification_NNFL/
│
├── data/
│   └── Emailscam.csv
│
├── notebooks/
│   └── email_classification.ipynb
│
├── models/
│   ├── best_rnn.weights.h5
│   └── best_lstm.weights.h5
│
├── results/
│   ├── confusion_matrix.png
│   ├── training_curves.png
│   └── classification_report.txt
│
├── README.md
└── requirements.txt

📊 Dataset

• Source: Mendeley Data
• Classes:
  • ham: legitimate messages
  • spam: unsolicited / fraudulent messages

Load dataset:

df = pd.read_csv(
    "https://raw.githubusercontent.com/ayeshatofa/Email-Classification-using-RNN/main/Emailscam.csv",
    encoding="latin1"
)

🔍 Exploratory Data Analysis (EDA)

The notebook includes visualizations for:

• Class distribution
• Word count analysis
• Punctuation frequency
• Stopword count
• URL frequency
• Frequent spam vs ham words

🛠 Text Preprocessing

Preprocessing steps include:

• Lowercasing
• Emoji removal
• Punctuation removal
• Stopword removal
• Stemming (PorterStemmer)
• Bigram generation
• Tokenization
• Sequence padding

processed_sentence = preprocessing(text)
processed_sentence = stopwordRemoval(processed_sentence)
processed_sentence = stem_text(processed_sentence)
processed_sentence = ' '.join(generate_ngrams(processed_sentence, n=2))

🤖 Model Development

SimpleRNN Model

• Embedding layer
• SimpleRNN(64)
• Dense layers with dropout

LSTM Model

• Embedding layer
• LSTM(64) with dropout
• Dense classifier

Training Setup

• Loss: binary_crossentropy
• Optimizer: adam
• Batch size: 32
• Epochs: 20
• EarlyStopping
• ModelCheckpoint

📈 Evaluation

RNN Performance

Metric	Score
Accuracy	0.9731
Loss	0.0949
Precision (weighted)	0.7492
Recall (weighted)	0.8656
F1 Score	0.8032

LSTM Performance

Metric	Score
Accuracy	0.9821
Loss	0.0705
Precision (weighted)	0.7492
Recall (weighted)	0.8656
F1 Score	0.8032

Confusion Matrix (Both Models)

[[483 0] [ 75 0]]

⚠ Note: Both models predicted all instances as ham due to dataset imbalance and heavy preprocessing, causing 0% recall on spam.

🧪 Custom Email Prediction

pred = model.predict(padded_seq)
pred_class = label_encoder.classes_[int(pred[0][0] > 0.5)]
print("Predicted Label:", pred_class)

⚠ Limitations

• Dataset imbalance → poor spam detection
• Heavy preprocessing removed useful patterns
• No class weighting used
• Simple RNN/LSTM architecture; lacks attention/transformers

🔮 Future Improvements

• Use class weights (e.g., class_weight='balanced')
• Apply SMOTE or oversampling for spam
• Include metadata features (sender, domain)
• Switch to BERT, DistilBERT, or Transformer models
• Refine preprocessing (keep URLs, moderate stopword removal)

▶ How to Run

1. Clone the repo:

git clone https://github.com/ayeshatofa/Email-Classification-using-RNN.git
cd Email-Classification-using-RNN

2. Install dependencies:

pip install -r requirements.txt

3. Open the notebook:

code CODE_2257_1002_1026.ipynb

4. Run all cells to train models / test predictions.