Case study: Up-sell recommender for Mall Group
Up-sell recommenders are crucial for all eCommerce companies, allowing their users to choose the best product for themselves. The up-sell recommenders show the alternative products to the viewed ones. Recently the alternatives have been calculated automatically by utilising machine learning methods and user behaviour data. Products are classed as similar to each other if they are often visited together.
Selecting alternatives to each product manually is not very sustainable in the case of a rapidly increasing number of products. This trend was accelerated immensely in Mall Group by the covid eCommerce boom and their strategy to become a marketplace. Still, a lot of customisation and parametrisation is needed for them to be able to consider sales business rules, agreements with vendors etc. Therefore, with our help, the client has decided to create their own recommender based on user behaviour data and compare it to the black-box recommenders using A/B testing.
Together with Mall Group, we built a modern up-sell recommender dealing with the following data:
- User behaviour data to consider similarities between products as the users perceive them (the product relation mode)
- Product content data (name, description etc.) to eliminate the cold-start problem (in case of new products have not sufficient visits)
- Sales data to maximise required KPIs
The product relation model pre-calculates the similarity of products based on the users’ behavioural history (user interactions with the products).
Using the item2vec approach, the products are projected (embedded) into the product relation space, where two products are near each other when they are treated similarly by users. Therefore, the output recommendations are the nearest items/products to the currently viewed product.