SMS Spam Classifier — LSTM

A deep-learning SMS spam detector built with TensorFlow/Keras (LSTM) and served via a Flask web app.

---

Project structure

spam_classifier/
├── app.py                    # Flask web app (trains model on startup, serves UI + /predict API)
├── LSTM_Spam_Detection.ipynb # Jupyter notebook with full EDA + training walkthrough
├── SMSSpamCollection.csv     # Dataset (tab-separated: label \t message)
└── requirements.txt

---

Setup

pip install -r requirements.txt

Place SMSSpamCollection.csv in the same folder as app.py.

---

Run the web app

python app.py

The script will:

1. Load and balance the dataset (down-sample ham to match spam count).
2. Tokenize and pad sequences.
3. Train a two-layer LSTM model (up to 30 epochs with early stopping).
4. Start a Flask server at 👉 http://127.0.0.1:5001.

Open the URL in your browser, type any SMS message, and click Classify Message.

---

Model architecture

Layer	Details
Embedding	vocab=500, dim=16, input_len=50
LSTM	units=20, dropout=0.2, return_seq=True
LSTM	units=20, dropout=0.2
Dense	1 unit, sigmoid activation

Loss: binary_crossentropy | Optimizer: adam

---

Dataset

The UCI SMS Spam Collection dataset — 5,572 messages labelled ham or spam.
Ham messages are down-sampled to 747 to match the spam count before training.

---

API

POST /predict

// Request
{ "message": "Congratulations! You've won a free prize. Call now!" }

// Response
{ "label": "SPAM", "score": 0.9821 }