Hashcoding Against Spam

Imagine: you naively posted your phone number in a huge Telegram channel called "Cats and Crypto," thinking the world was full of kindness. But now your messenger is a hellish mix of:

• "Congratulations! You've won 10 kg of buckwheat! Send us your card number!"
• "Your account has been hacked. Urgently send a photo of your passport and your favorite cat!"
• "Exclusive: how to make money breeding snails in the garage. Webinar starts in 5 minutes!"

Instead of running off to the wilderness, you decide to teach a machine to tell spam (crypto scams) from ham (normal messages like "Mom, buy bread"). That's exactly what this project does.

What's Inside

A binary text classifier written in pure C++. No external ML libraries. Input: raw text. Output: spam or ham label.

How It Works

1. Text vectorization. Each message is split into tokens and lowercased. Every token is hashed via MurmurHash3; the hash is used as an index in a fixed-size vector (Feature Hashing / Hashing Trick). The vector is L2-normalized.

2. Training. Logistic regression is trained with full-batch gradient descent and Cross-Entropy Loss. The learning rate decays on a schedule: 0.01 → 0.005 (at epoch 400) → 0.001 (at epoch 4000). Every 10 epochs the model is evaluated on validation, and the weights with the best metrics are saved (early stopping).

3. Class weighting. Before training, class weights are computed based on the distribution in the training set — so the model doesn't collapse into predicting the majority class.

4. Threshold tuning. After training, thresholds from 0.1 to 0.99 are swept with a step of 0.01. The threshold with the highest Accuracy is chosen, subject to Recall >= 0.6.

Tech Stack

• C++11, STL
• MurmurHash3 (custom implementation), L2 normalization
• Logistic regression, gradient descent
• Feature Hashing

Project Structure

• MurmurHash3.cpp — MurmurHash3 implementation (32-bit, bitwise operations, rotations, final mixing).
• hashing.cpp — text vectorization (text_to_features), LogisticRegression class (training, prediction, evaluation, threshold tuning), CSV parsing.
• main.cpp — data loading, class weight computation, training launch, metrics output.
• my_tests.cpp — tests: hash determinism and avalanche effect, L2 normalization, model convergence on synthetic data.

Input Data

CSV format class,text, first line is the header:

class,text
spam,Free money now!!!
ham,Hi, how are you?

• entrypoint/data_train.csv — training dataset.
• entrypoint/dataset.csv — validation dataset.

Metrics

Target metrics on validation:

• Accuracy >= 0.6
• Recall >= 0.6

Build and Run

g++ -O3 -std=c++11 main.cpp -o spam_filter
./spam_filter

Tests

g++ -O3 -std=c++11 my_tests.cpp -o tests
./tests

Three test groups:

• hash_test — identical inputs produce identical hashes, different inputs produce different hashes.
• normalization_test — L2 norm of text_to_features output equals 1.
• Log_reg_test — the model separates two synthetic classes.

---

P.S. The author is not responsible if, after studying the code, you start seeing hash vectors in your nightmares.