BlendCAL — Web-Session Conversion Prediction

Overview

BlendCAL is an end‑to‑end ML project for predicting whether a web session will convert.
The pipeline covers: feature engineering, an ensemble of CatBoost + XGBoost + LightGBM with isotonic calibration, a FastAPI inference service, a Streamlit UI, an Airflow DAG for orchestration, and full Docker setup.

Course project (Skillbox, ML specialization).

---

Data volume

• ga_sessions.csv — 1,860,042 rows
• ga_hits.csv — 15,726,470 rows

---

Project structure

final_project/
├─ dags/                          # Airflow DAG (blendcal_inference.py)
├─ project/
│  ├─ modules/                    # ETL & feature pipeline
│  │  ├─ extract_csv.py
│  │  ├─ prepare.py
│  │  └─ ensemble.py
│  ├─ data/                       # raw / landing / staging / predictions
│  └─ artifacts/                  # prep_params.json, freq_maps.json, models, calibration
├─ api/app/                       # FastAPI: main.py, artifacts_loader.py, preprocessor.py
├─ app/                           # Streamlit UI (streamlit_app.py)
├─ docker_airflow/                # Dockerfile + docker-compose for Airflow
├─ docker-compose.yml             # API + UI
├─ requirements-api.txt
├─ requirements-ui.txt
├─ MODEL_INFO.json
├─ VERSION
└─ docker_airflow/README_airflow.md

---

Tech stack

• ML: CatBoost, XGBoost, LightGBM (weighted ensemble + isotonic calibration)
• Preprocessing: median imputation, quantile clipping, frequency encoding, cyclic time features (sin/cos)
• API: FastAPI, Pydantic, Uvicorn
• UI: Streamlit
• Orchestration: Airflow (PythonOperator, Docker stack)
• Containerization: Docker, docker‑compose

---

Quickstart

1) API + UI (Docker)

docker compose up --build

• FastAPI Swagger: http://localhost:8000/docs
• Streamlit: http://localhost:8501

2) Airflow (Docker)

See docker_airflow/README_airflow.md for details. TL;DR:

cd docker_airflow
docker compose down -v
docker compose up airflow-init
docker compose up -d webserver scheduler

• Airflow UI: http://localhost:8080 (admin / admin)

---

Results

• ROC‑AUC: 0.86
• F1 macro: 0.75
• Holdout period: 2021‑11 → 2021‑12

Model metadata & artifacts are recorded in MODEL_INFO.json.

---

Useful links

• Airflow how‑to: docker_airflow/README_airflow.md
• Model passport: MODEL_INFO.json

---

Author

• Konstantin Nikiforov — Skillbox ML specialization (2025)

License

This project is licensed under the MIT License. See LICENSE for details.