Product Matching Engine 🚀

High-precision E-commerce deduplication system based on CV and NLP.

7th Solo Place out of 110 teams in the Ozon Tech ML Competition. ROC-AUC Score: 0.9216

---

📌 Overview

This repository contains a production-ready solution for the product matching task. The engine predicts whether two products are identical (target = 1) or different (target = 0) by analyzing multi-modal data: titles, technical attributes, and images.

🏁 Key Results

Model	ROC-AUC	Focus
AutoGluon (Stacked)	0.9216	Maximum accuracy (Top-7 Leaderboard)
HistGradientBoosting	0.9100	Production-ready, lightweight & fast

📦 Data Architecture

The system processes five distinct data streams provided by Ozon Tech:

• Textual: Titles, descriptions, and pre-trained BERT embeddings.
• Visual: ResNet embeddings of product images.
• Technical: Product categories and structured attributes (attributes.csv).
• Relational: Pairs of products for training (train.csv).

🔧 Engineering Pipeline

The core strength of this solution lies in advanced feature engineering:

• NLP Features: Levenshtein distance, Jaccard similarity, string length analysis.
• Attribute Analysis: Jaccard similarity of attribute sets, count differences, binary matching flags.
• CV Analysis: Cosine and Euclidean distances between image embeddings, entropy-based features.
• Categorical: Multi-level subcategory matching.

🧠 Model Architectures

1. AutoGluon Tabular (The Heavy Hitter)

A sophisticated multi-model stack with automatic ensembling.

• Presets: best_quality for maximum precision.
• Strategy: Automated hyperparameter tuning and multi-layer stacking.
• File: notebooks/04_AutoML_Training_AutoGluon.ipynb

2. HistGradientBoosting + Optuna (The Fast Track)

An alternative lightweight pipeline designed for high-load environments.

• Optimization: Bayesian search via Optuna.
• Strengths: Handles class imbalance, minimal inference latency.
• File: notebooks/03_Model_Training_HGB.ipynb

📂 Repository Structure

Product-Matching-Engine/
├── assets/             # Project branding and images
├── notebooks/          # Research & Development
│   ├── 01_Feature_Extraction.ipynb
│   ├── 02_Category_Analysis.ipynb
│   ├── 03_Model_Training_HGB.ipynb
│   └── 04_AutoML_Training_AutoGluon.ipynb
├── src/                # Production scripts (Inference)
│   ├── inference_hgb.py
│   └── inference_autogluon.py
└── requirements.txt    # Project dependencies

🔍 Analysis & Future Work

The high performance was achieved through massive feature generation and AutoML. Analysis of top-3 solutions shows that further gains can be made by fine-tuning Transformers (BERT) specifically on product categories to capture deeper semantic relationships.

---

Developed by Nikolay Alymov Expertise in AI, RecSys, and Enterprise Automation.