System for personalized recommendations on the Kazan Express website
Project information
Client
Kazan express is a marketplace for Chinese and Russian good with free one day shipping
Has been working since 2017. The fastest growing marketplace in Europe.
Increase the number of purchases
Task
KazanExpress.ru website users
Audience
Project concept
The customer sees recommended products that they may want to purchase.
Development process
Step 01
Approval
Step 02
Data preparation
Step 03
Architectural design
Step 06
User testing (KazanExpress employees)
Step 05
Train the model
Step 04
Development
How does the system work
User visits client's site
Selects and product and places it in their basket
The system offers products that the customer may need
Average check value rises
What the recommendations are based on
Purchase history of a specific user
Comparing purchases of different users
We look at the customer's «wish list» and previous
We do not take into account:
location and gender.
Two users have similar purchases - we call them «twin users».
We recommend what the «twin user» has already bought.
Technologies
Spark library
framework for processing large data streams
ALS matrix decomposition
an algorithm that processes user actions in real time
intentionally limited the system for users who have made at least four purchases
Make bi-weekly system updates to prevent a system overload
A marketplace is a huge array of data (several million rows in tables).
That's why we:
That's why we:
Difficulties faced
Problem
Due to no clear quality criteria, acceptance of the project was delayed
Output
It is necessary that during negotiations with the client that a clear understanding of the quality of the metercs and means of aquiring them is established
Problem
At times the service worked slower than we wanted
Decision
We found a hard-to-reproduce bug in the code and made additional optimizations to the speed of the algorithms.
Quote
«The only complaints we received during testing was that a man could have a pink phone case recommended, however we resolved that issue».
Nadezhda
Manager
Project team
Manager
Nadezhda Zagvozkina
Data Scientist
Oleg Durygin
Junior Data Scientist
Александра Щетинина
What we have learned
Work more thoughtfully with recommender libraries
Handle huge amounts of data competently
Our plans
We are negotiating with the client for support and improvement:
new parameters (gender, age, location)
optimization
Nadezhda
«We are planning to further develop in recommendations and build up competencies».
Manager
Where the implementation can be applied
Marketplaces, online stores and, in general, any entity that sells goods.
However recommender systems work best when given large data sets
Contact us
PRODUCTS
SECTIONS
STUDIOS
Russia, Innopolis, st. Universitetskaya, 7
2023 © EORA. All right reserved