Digital illustration of computer monitor displaying high heels, with images of similar shoes emerging from the screen.

Visual Recommendation System

Image-based recommendation system which recommends products based on their visual similarity.

A hand holding a smartphone with a picture of a pink bag on the screen. In the background - clothes and accessories: shoes, lipstick, bag (the same as on the phone screen), sweater and glasses. Everything in rose-purple colors.

How does it work?

We are using transfer learning from the field of image classification to extract features from new data.

What is transfer learning?

Transfer learning involves applying knowledge from one field in another field. In our recommendation system, we are using an already trained model and training only the top layers to adapt to new cases.
  • 01


    There is no need for a priori data thanks to transfer learning. The only required component from stores is a database of the product images to feed the model, so it's possible to build a working demo of this recommendation system for your store within a day.

  • 02

    Development opportunities

    Adding filtering based on attributes would enhance the performance of the system. For example if we filter by category, we can recommend products for men and women separately from the category. This way, a men’s t-shirt with flowers will not be recommended as a similar product for a woman viewing t-shirts with flowers for women. Product can be improved in the future given feedback is collected based on user clicks on the website. Items clicked from recommendation would be considered a successfully recommended item and therefore the system could learn from user engagement with the given recommendation.

  • 03


    This system can be used to recommend visually similar products when looking for alternatives to the currently viewed item. It's perfect for a starting point, a so called cold start, in behavioral recommendations when there is not sufficient data gathered yet to suggest items based on user's search history.