Edit Content
We Are Akira

Deep Learning Recommender System


To pivot business operations from store focused to customer focused by using upsell and cross-selling as strategies for better traction online

Business Problem

Big Bazaar wants to strengthen their online shopping platform shop.bb by improving their online shopping site to attract new customers and to convert the existing brick and mortar customers to Online/Omni-channel customers

Solution overview

  • The recommendation engine has been developed using a custom Learning to Rank (LTR) model. This model learns user and item embeddings to search relevant items for a customer. Online transaction data is the primary source to train the model. Customer and item metadata are also used to enhance the accuracy. [PD- suggest removing The RecNet model is very fast and easy to train, works well even with low volume of data and has shown 80 – 85% accuracy on withheld test sets]
  • The technical implementation of the recommender system leverages AWS EC2 as a platform for orchestration, and the model implements Tensorflow and Keras for ranking. The model is trained during each scheduled session and the outputs are pushed to a S3 bucket. The online purchase platform fetches relevant information from this location as recommendations and popular items.

Solution Flow

Business outcome

Triggers such as these have been seen to be successful in controlled experiments run by the data science team at AKIRA which have  (Upselling product pairs and similar customers’ cart additions) contributed to a significant revenue uplifts to the tune of 3%

Solution Architecture

Tool Stack

  • AWS EC2 and Docker – The entire application is wrapped in a container and scheduled with an AWS EC2 instance.Container helps the application to be migrated or used anywhere.
  • S3 – AWS Simple Storage Service the primary source for input pipelines as well as destination for inference pipelines.Serves as a backend for the online platform to fetch inference.
  • Tensorflow and Keras – The model is completely built from the ground up using Tensorflow and Keras framework. This makes it very fast and easy to customize.