SMS Spam Detection Using Neural Networks
Nathanael Mbale
SMS Spam Detection Using Neural Networks
Quick Overview
Problem: Build a binary text classifier to distinguish between legitimate messages ("ham") and spam SMS using natural language processing and deep learning.
Solution: Implemented a neural network with text vectorization and word embeddings to classify SMS messages, achieving robust spam detection through sequential text processing.
Impact: Created a production-ready spam filter with high accuracy on unseen messages, demonstrating proficiency in NLP, text preprocessing, and TensorFlow text classification pipelines.
Step-by-Step Implementation
Step 1: Environment Setup & Library Installation
What: Installed a stable TensorFlow version and imported required libraries.
Key Libraries:
TensorFlow / Keras for neural networks
Pandas for data handling
TensorFlow Datasets for utilities
NumPy for numerical operations
Matplotlib for visualization
Version Control: Ensured a stable TensorFlow installation by removing
tf-nightly if present.Step 2: Data Acquisition
What: Downloaded the SMS Spam Collection dataset using
wget.Dataset Files:
train-data.tsv: Training messagesvalid-data.tsv: Testing and validation messagesFormat: Tab-separated values (TSV) with label and message columns.
Source: FreeCodeCamp project data repository.
Step 3: Data Loading & Exploration
What: Loaded TSV files into Pandas DataFrames.
Structure:
Column 1: Label (
ham or spam)Column 2: Message text content
Initial Analysis: Examined the first few rows to understand data format and distribution.
Key Finding: Dataset is imbalanced, with significantly more ham than spam messages.
Step 4: Label Encoding
Problem: The model requires numerical labels, but the dataset contains text labels.
Solution: Binary encoding of labels.
ham → 0 (legitimate message)spam → 1 (spam message)Implementation: Converted labels into NumPy arrays.
Result: Training and validation labels represented in binary format (0/1).
Step 5: Text Preprocessing Pipeline
Custom Standardization Function: Implemented a text cleaning pipeline.
Converted text to lowercase
Removed HTML tags such as
<br /> using regular expressionsRemoved punctuation and special characters
Purpose: Normalize text variations (for example, "FREE" versus "free") for consistent processing.
Impact: Reduced vocabulary size and improved model generalization.
Step 6: Text Vectorization Configuration
Layer:
TextVectorization from Keras.Parameters:
max_tokens = 10,000: Vocabulary size limitsequence_length = 120: Fixed output length with padding and truncationoutput_mode = 'int': Convert words to integer indicesstandardize = custom_standardization: Apply text cleaning functionVocabulary Building: Adapted the vectorization layer on training messages only.
Result: Learned a word-to-index mapping based on the training corpus.
Step 7: Message Vectorization
Process: Converted text messages into integer sequences.
Training data:
train_messages → train_sequencesTest data:
test_messages → test_sequencesExample Output:
Original:
"Free money now!"Vectorized:
[42, 187, 933, 0, 0, ...] (padded to length 120)Benefit: Enabled neural network processing of numerical sequences.
Step 8: Neural Network Architecture Design
Model Type: Sequential neural network for binary text classification.
Architecture:
Embedding layer: Converts word indices into 16-dimensional dense vectors
Dropout layer (20%) for regularization
GlobalAveragePooling1D layer to aggregate sequence information
Dropout layer (20%) before final output
Dense output layer with 1 neuron and sigmoid activation
Why This Architecture:
Embeddings learn semantic word representations
Global average pooling handles variable-length sequences
Dropout reduces overfitting
Sigmoid outputs probabilities between 0 (ham) and 1 (spam)
Step 9: Model Compilation
Optimizer: Adam, suitable for NLP tasks due to adaptive learning rates.
Loss Function: Binary crossentropy, standard for binary classification.
Metric: Accuracy as the primary evaluation metric.
Configuration: Optimized for binary text classification.
Step 10: Dataset Analysis
Class Distribution Check:
Counted ham versus spam messages in the training set
Calculated spam ratio percentage
Purpose: Identify class imbalance, which affects evaluation and interpretation.
Finding: Dataset contains more ham messages than spam.
Implication: Accuracy must be interpreted carefully due to imbalance.
Step 11: Model Training
Training Configuration:
Epochs: 60
Validation data: Vectorized test sequences
Verbose: 1 (progress bars enabled)
Data Used: Vectorized message sequences and binary labels.
Monitoring: Tracked accuracy and loss on both training and validation sets.
Result: Model learned patterns distinguishing spam from legitimate messages.
Step 12: Post-Training Verification
Immediate Testing: Generated predictions for a sample spam message.
Example:
"you have won £1000 cash! call to claim your prize."
Vocabulary Check: Verified that 10,000 words were learned.
Sanity Check: Confirmed output probabilities aligned with expectations.
Purpose: Ensure the model learned meaningful spam indicators such as "won," "prize," and "cash."
Step 13: Model Evaluation
Test Set Performance: Evaluated on unseen validation data.
Metrics Reported:
Test accuracy
Test loss (binary crossentropy)
Purpose: Measure generalization to new messages.
Result: High accuracy indicating effective spam detection.
Step 14: Model Persistence
What: Saved the trained model to disk.
Format: Keras native format (
.keras).Filename:
spam_model.kerasBenefit: Enables model reuse without retraining.
Use Case: Deployment or continued development.
Step 15: Prediction Function Implementation
Function:
predict_message(pred_text)Input: Single message string.
Process:
Vectorize input text using the trained
vectorize_layerGenerate a probability prediction between 0 and 1
Apply threshold: ≥ 0.5 → spam, < 0.5 → ham
Output Format:
[probability_float, label_string]Example:
[0.9234, 'spam'] or [0.1567, 'ham']Key Design: Returns both probability and label for interpretability.
Step 16: Comprehensive Testing
Test Suite: Seven diverse messages covering edge cases.
Examples Include:
Normal conversation
Obvious spam
Scheduling messages
Service notifications
Prize scams
Casual reminders
Personal stories
Validation: Compared predictions against expected labels.
Output: Detailed report with probability, prediction, and correctness.
Result:
✅ Successfully classified all test cases.
Step 17: Final Validation
Automated Test: Built-in
test_predictions() function.Pass Criteria: All seven test messages correctly classified.
Result:
✅ "You passed the challenge. Great job!"
Verification: Model generalizes effectively across varied message types.
Like this project
Posted Dec 27, 2025
Built an SMS spam detector using deep learning and NLP techniques.
Likes
0
Views
0
Timeline
Dec 20, 2025 - Dec 23, 2025
