CWV-Predict: Core Web Vitals Prediction from Page Features

Type: Academic | Domain: SEO, Web Performance
Hugging Face: syeedalireza/cwv-predict

Predict LCP, INP/FID, and CLS from page and server features before deployment (pre-deployment CWV estimation).

Author

Alireza Aminzadeh

• Hugging Face: syeedalireza
• LinkedIn: alirezaaminzadeh
• Email: alireza.aminzadeh@hotmail.com

Problem

Core Web Vitals (CWV) affect rankings and UX. Estimating CWV before deployment helps prioritize fixes and avoid regressions.

Approach

• Features: Resource counts, DOM size, script size, server response time, image dimensions, etc.
• Targets: LCP (ms), INP/FID (ms), CLS (score).
• Models: Gradient boosting (XGBoost/LightGBM) for tabular features; optional lightweight embeddings for URL/path.

Tech Stack

Category	Tools
ML	scikit-learn, XGBoost, LightGBM
Data	pandas, NumPy
Evaluation	sklearn metrics, optional MLflow

Setup

python -m venv .venv
.venv\Scripts\activate   # Windows
pip install -r requirements.txt

Usage

# Train (expects data in data/ or env DATA_PATH)
python train.py

# Predict on new samples
python inference.py --input data/sample.csv --output predictions.csv

Project structure

01_cwv-predict/
├── config.py          # Paths and column names
├── train.py           # Train XGBoost regressors per CWV metric
├── inference.py       # Batch prediction from CSV
├── requirements.txt
├── .env.example
├── data/
│   ├── cwv_features.csv   # Sample training data (features + targets)
│   └── sample.csv         # Sample input for inference (features only)
└── models/                # Saved models and metrics (after train)

Data

• Sample data (included): data/cwv_features.csv (features + lcp_ms, inp_ms, cls_score), data/sample.csv (features only for inference).
• Columns: dom_size, resource_count, script_bytes, image_count, server_ttfb_ms; targets: lcp_ms, inp_ms, cls_score.
• Override paths via DATA_PATH in .env. See .env.example.

License

MIT.