Portfolio

Automation of product placement on OZON

More than 20,000 stores sell goods on the Ozon marketplace, and some of those stores offer hundreds of thousands of items, making managing product placement challenging. To simplify the process of listing new products, Ozon sought our expertise to create a tailored solution.

Challenges in Publishing Products on Ozon

Strict requirements for product cards.
Complex category system with over 4,800 categories, among which the seller must choose only one.
The interface isn’t optimized for international users, complicating the process of finding the correct categories and properties.
Manual placement of a large volume of items is too time-consuming.

The Task

We were required to fully automate the process of filling in cards and uploading goods to Ozon. This solution would:

Eliminate the need for manual data entry.
Significantly accelerate the listing process.

Our Solution

We developed a B2B product management system powered by machine learning, designed to act as a bridge between sellers and the marketplace. Its main functions are:

Data Conversion: Transforms seller-provided data into Ozon-compliant formats.
Automated Product Cards: Creates ready-to-publish product listings.
Real-Time Updates: Tracks and updates product information as needed.

How it works

Sellers upload a file with their product details and link their Ozon accounts.
Using machine learning, the system identifies the appropriate category for each product.
A neural network standardizes product properties and adapts them to Ozon’s format.
Listings are automatically published on Ozon.

Product Classification

The system highlights key characteristics of a product by analyzing its description, photos, and category. This process is applied to all products in Ozon's categories.

When a seller adds a new product, it is automatically assigned to the Ozon category containing the most similar items. This approach eliminates the need to retrain the model when the platform's category structure changes, reducing both costs and associated risks.

Property Matching

Products often have synonymous terms for materials or features. For instance, a seller might list "leatherette," while Ozon uses "artificial leather."

We aim to use machine learning to determine that they are the same thing. The system achieves this by analyzing the linguistic 'proximity' of properties and their meanings.

Automated Naming

The system generates standardized product names, following a structured format:

Type + Brand + Model + Key Features.

Manual Moderation of the Content

Quality control is essential, as the system is not immune to errors. At each stage of the process, a moderator manually reviews the model's predictions, verifies category assignments, and validates product properties. If the moderator makes corrections, corrects errors, or rejects a product, the system incorporates this feedback to improve its accuracy and performance.

Continuous Updates

Once the product is published on Ozon, it remains under continuous monitoring. If the seller updates the product information, the system automatically detects these changes and updates the product description on Ozon in real time.

Conclusion

We developed a user-friendly B2B product management system powered by machine learning, acting as an interface between sellers and the marketplace. The system:

standardizes seller data to align with Ozon's requirements,
automatically generates ready-to-use product listings, and
ensures quick and error-free publication of goods.

Results

We developed a system that automatically categorizes products and converts their attributes into Ozon's format.
Sellers no longer waste time filling out cards manually.
Publishing hundreds of thousands of products now takes just a few days.
Product descriptions are more accurate and customer-focused, resulting in improved sales performance.
Using the Epoch8' system, over 500,000 products have already been successfully listed on Ozon.

2024-11-25 14:29