VusalShahbazovAz/smishing-AZ-SC

SMS Classification Dataset - Azerbaijan

Citation

If you use this dataset in your research, please cite:

@article{shahbazov2026azerbaijani,
  title     = {SMS dataset for multi-class classification of ham, spam, and smishing in Azerbaijani language},
  author    = {Vusal Shahbazov},
  journal   = {Problems of Information Technology},
  volume    = {17},
  number    = {1},
  pages     = {32--39},
  year      = {2026},
  issn      = {2304-0157},
  doi       = {10.25045/jpit.v17.i1.04},
  url       = {https://jpit.az/en/journals/355}
}

This project evaluates multiple machine learning algorithms for SMS classification on an Azerbaijani language dataset. The dataset contains SMS messages classified into three categories: ham, spam, and smishing.

Github

https://github.com/vusalshahbaz/sms-classification-dataset-azerbaijan

Dataset

• Source file: dataset.csv
• Total entries: 4,529
• Labels:
  • ham: 3,030 (66.90%)
  • smishing: 612 (13.51%)
  • spam: 887 (19.58%)
• Test size: 20% (default)
• Random state: 42

Dataset Statistics

Generate comprehensive statistics about the dataset:

python dataset_statistics.py --dataset dataset.csv

This will show:

• Total number of messages
• Class distribution
• Average words, characters, and sentences per message
• Messages containing phone numbers (obfuscated)
• Messages containing emails
• Messages containing URLs
• Statistics by class
• Sender statistics

Preprocessing

The dataset undergoes Azerbaijani character normalization to unify text representation:

• ə/Ə → e/E
• ı/I/İ → i/I
• ç/Ç → c/C
• ğ/Ğ → g/G
• ö/Ö → o/O
• ü/Ü → u/U
• ş/Ş → s/S

Feature Extraction

Most models use TF-IDF vectorization with the following parameters:

• ngram_range: (1, 2) - unigrams and bigrams
• min_df: 2 - minimum document frequency

CNN uses a different approach:

• Text tokenization with word-level tokenization
• Sequence padding to fixed length (max_len=100)
• Embedding layer for word representations

Algorithms and Hyperparameters

The following table lists all algorithms used in this project along with their hyperparameters and suggested ranges for tuning:

Method	Hyperparameters	Range
Logistic Regression	max_iter, C (inverse regularization)	100, 500, 1000 - 0.01, 0.1, 1.0, 10.0
Linear SVM	C (regularization), max_iter	0.01, 0.1, 1.0, 10.0, 100.0 - 1000, 2000, 5000
Passive Aggressive	C (regularization), max_iter	0.01, 0.1, 1.0, 10.0 - 100, 500, 1000
Multinomial NB	alpha (smoothing)	0.1, 0.5, 1.0, 2.0
Decision Tree	max_depth, min_samples_split, min_samples_leaf	5, 10, 20, 30, None - 2, 5, 10 - 1, 2, 4
Random Forest	n_estimators, max_depth, min_samples_split	50, 100, 200, 500 - 5, 10, 20, None - 2, 5, 10
KNN	n_neighbors, weights, metric	3, 5, 7, 9, 11 - uniform, distance - euclidean, manhattan, cosine
CNN	max_words, max_len, embedding_dim, filters, kernel_size, dense_units, epochs, batch_size	5000, 10000, 20000 - 50, 100, 150 - 64, 128, 256 - 64, 128, 256 - 3, 5, 7 - 32, 64, 128 - 5, 10, 20 - 16, 32, 64

Current Model Configurations

Logistic Regression

• max_iter: 1000
• solver: lbfgs
• C: 10.0 (optimized)

Linear SVM

• C: 1.0 (optimized)
• max_iter: 2000 (optimized)

Passive Aggressive

• max_iter: 1000
• C: 0.01 (optimized)

Multinomial Naive Bayes

• alpha: 0.5 (optimized)

Decision Tree

• random_state: 42
• max_depth: default (None)
• min_samples_split: default (2)
• min_samples_leaf: default (1)

Random Forest

• n_estimators: 100
• random_state: 42
• max_depth: default (None)
• min_samples_split: default (2)

K-Nearest Neighbors (KNN)

• n_neighbors: 5 (optimized)
• weights: default (uniform)
• metric: default (minkowski/euclidean)

Convolutional Neural Network (CNN)

• max_words: 10000
• max_len: 100
• embedding_dim: 256 (optimized)
• filters: 128
• kernel_size: 3
• dense_units: 64
• epochs: 20 (with early stopping)
• batch_size: 32
• random_state: 42
• dropout: 0.5
• Features: Learning rate scheduling, early stopping with patience=5

Recurrent Neural Network (RNN)

• max_words: 10000
• max_len: 100
• embedding_dim: 256 (optimized)
• rnn_units: 256 (optimized)
• dense_units: 128 (optimized)
• dropout_rate: 0.3 (optimized)
• epochs: 30 (with early stopping)
• batch_size: 32
• random_state: 42
• Features: Learning rate scheduling, early stopping with patience=7

Usage

Prerequisites

Install required packages:

pip install pandas scikit-learn numpy

For CNN and RNN models (optional, requires TensorFlow):

pip install tensorflow

Note: TensorFlow requires Python 3.8-3.12. If TensorFlow is not available, the script will run all other models and skip CNN/RNN.

Running the Evaluation Script

Run the evaluation script with default settings:

python evaluate_models.py

With custom arguments:

python evaluate_models.py --dataset dataset.csv --test-size 0.2 --random-state 42

Arguments:

• --dataset: Path to the CSV dataset (default: dataset.csv)
• --test-size: Fraction of dataset for testing (default: 0.2)
• --random-state: Random seed for reproducibility (default: 42)

Running Dataset Statistics

Generate comprehensive dataset statistics:

python dataset_statistics.py --dataset dataset.csv

Save statistics to JSON file:

python dataset_statistics.py --dataset dataset.csv --output stats.json

Evaluation Metrics

All models are evaluated using macro-averaged metrics:

• Accuracy: Overall classification accuracy
• Recall: Macro-averaged recall across all classes
• Precision: Macro-averaged precision across all classes
• F1 Score: Macro-averaged F1 score across all classes

Results

Model Performance (Latest Run)

Algorithm	Accuracy	Recall	Precision	F1 Score
Passive Aggressive	0.9338	0.8788	0.8857	0.8821
Linear SVM	0.9327	0.8718	0.8861	0.8787
Logistic Regression	0.9294	0.8649	0.8830	0.8736
CNN	0.9205	0.8508	0.8734	0.8614
Multinomial NB	0.9172	0.8560	0.8610	0.8576
Random Forest	0.9051	0.8220	0.8466	0.8334
Decision Tree	0.8786	0.7931	0.7948	0.7939
KNN	0.7152	0.4343	0.7678	0.4441
RNN	0.6689	0.3333	0.2230	0.2672

Performance Summary

• Best Overall Accuracy: Passive Aggressive (93.38%)
• Best F1 Score: Passive Aggressive (0.8821)
• Best Recall: Passive Aggressive (0.8788)
• Best Precision: Linear SVM (0.8861)

Notes

• Results are based on a 20% test split with random_state=42
• All metrics are macro-averaged across classes (ham, spam, smishing)
• CNN and RNN models use optimized hyperparameters with learning rate scheduling
• RNN performance may improve with further hyperparameter tuning

Hyperparameter Tuning

Individual model hyperparameter tuning scripts are available in the methods/ directory. Each method has its own tuning script and best configuration documented.

Example: Tuning CNN

cd methods/cnn
source ../../venv_tf/bin/activate
python tune.py

Project Structure

.
├── dataset.csv                    # Main dataset file
├── evaluate_models.py            # Main evaluation script
├── dataset_statistics.py         # Dataset statistics generator
├── utils.py                      # Shared utilities
├── methods/                      # Individual method implementations
│   ├── cnn/
│   ├── rnn/
│   ├── logistic_regression/
│   └── ...
└── README.md                     # This file