Our client is a leading recommendation system vendor to online apparel retailers. Their existing system had plateaued in performance — customer profile data was not adding value to recommendation accuracy and Item features were not available. They needed to develop the next generation of recommendation technology to increase customer engagement levels.


Using a recommendation system based on a single model has its own limitations so, we decided to build multiple base models and combine them to build a hybrid recommendation model that also takes into account customer profiles and the type of products.


  • The Tatras team built multiple base recommenders using traditional approaches as well as novel item and customer profiles.
  • Approaches were developed to scale base recommender scores to remove implicit scale bias.
  • Product sales behaviour as well as consumer purchase behaviour were clustered to extract new features such as “eclectic customers”, and “ephemeral items”.
  • Automated tagging of products using deep learning enabled improved incorporation of product data within recommendations.
  • Items were modeled with respect to their sales patterns and features extracted from the same.
  • Customers were modeled based on the types of products they had purchased over time and when they were purchased within the product’s life cycle.
  • Local rather than global weights were learnt to create ensembles of recommenders for clusters of users.


Feature Engineering, Scaling, Collaborative Filtering, Random Forests, Ensembles, Clustering, Bayesian Mixture Models, EM, Convolutional Neural Networks, Transfer Learning, Variational Auto encoders.


We were able to improve customer engagement and were also able to drive up sales and increase conversions rate.