Domain Name Classification & Price Prediction

Machine Learning Pipeline for Domain Valuation using Structural and Linguistic Features

ML-based domain name classification across 5 pricing categories using feature engineering and ensembling techniques.

---

Overview

This project implements a machine learning pipeline to predict domain name pricing categories based on structural and linguistic features. Using 8.7 million domain dataset, the system extracts 16 engineered features and applies three classification algorithms with hyperparameter optimisation.

The Result? Random Forest achieved 65% accuracy on 5-class imbalanced dataset, with +4% improvement through systematic hyperparameter tuning.

---

Key Results

Model Performance Comparison

Model	Default Accuracy	Tuned Accuracy	Improvement
Logistic Regression	56.0%	58.0%	+2.0%
Decision Tree	59.0%	62.0%	+3.0%
Random Forest	61.0%	65.0%	+4.0%

Best Model: Random Forest (Tuned)

Optimal Hyperparameters:

• n_estimators=200 (trees in ensemble)
• max_depth=20 (tree depth)
• min_samples_split=2 (split threshold)

Per-Category Performance (F1-Scores):

• Category 1 (Expensive): 0.96
• Category 2: 0.47
• Category 3: 0.37
• Category 4: 0.29
• Category 5 (Cheap): 0.22

---

Implementation

Three-Stage Pipeline:

Stage 1: Feature Engineering

• Extracted 16 features from domain structure (length, TLD, entropy, character composition).
• Selected top 10 features via correlation analysis.
• Applied StandardScaler normalisation.

Stage 2: Baseline Models

• Trained 3 algorithms with default parameters.
• Evaluated on stratified 70-30 train-test split.
• Identified Random Forest as best performer (61%).

Stage 3: Hyperparameter Optimisation

• Grid Search with 3-fold cross-validation.
• Tested 54 parameter combinations (Random Forest).
• Final accuracy: 65% (+4% improvement).

---

Project Highlights

• Massive Scale: Handled 8.7M domain dataset with efficient sampling strategy.
• Feature Engineering: 16 structural & linguistic features (TLD, entropy, character patterns).
• Systematic Optimisation: GridSearchCV across 3 algorithms with CV.
• Measurable Gains: +4% accuracy improvement through tuning.
• Imbalance Handling: Stratified sampling maintains class distribution.
• Visualisation: Confusion matrices and performance comparisons.
• Reproducible: Complete pipeline with random_state set throughout.

---

Tech Stack

• Algorithms: Logistic Regression, Decision Tree, Random Forest.
• Feature Engineering: TLD extraction (tldextract), Shannon entropy, character analysis.
• Optimisation: GridSearchCV with 3-fold CV.
• Evaluation: Precision, Recall, F1-Score, Confusion Matrices.
• Dataset: 8.7M domains across 5 pricing categories.
• Framework: scikit-learn, Pandas, NumPy, Matplotlib, Seaborn.

---

Limitations & Challenges

Class Imbalance

• Issue: Category 1 (57% of data) vs Category 5 (2% of data).
• Impact: Model biased toward expensive domains (Cat 1: 96% F1 vs Cat 5: 22% F1).
• Current Mitigation: Stratified sampling, per-class metrics.
• Future Improvement: SMOTE oversampling, class_weight='balanced'.

Computational Constraints

• Issue: 8.7M rows require significant compute time.
• Current Approach: 10% random sampling (609K samples) for training.
• Impact: Potential 2-4% accuracy loss vs full dataset.
• Future Improvement: Train on full dataset with cloud GPU resources.

Feature Limitations

• Issue: Only structural features (no domain age, backlinks, traffic data).
• Impact: Missing key pricing indicators.
• Future Improvement: Integrate external domain metadata.

Important Notes

Data Requirements

This project uses a domain pricing dataset with 8.7M rows. A 500K sample dataset is provided in this repository: sample_domains.csv

Dataset format:

• Columns: Domain (string), category (1-5)
• Categories: 1=expensive to 5=cheap
• Maintains original 95.6% class imbalance

Installation & Usage

pip install -r requirements.txt