Recommender System in E-commerce

Challenge

For a good user experience, customer-specific content on the website is of utmost importance. Our client wanted to enhance their online shop by providing personalized product recommendations. The development of recommender systems is a central component, as they can recognize purchasing patterns along the purchase history. In this case, most users log in only after shopping, so customer master data cannot be used to create product recommendations. Additionally, new products are regularly introduced in the online shop, meaning there is no sufficient sales history for these products (Cold Start Problem). Our client faced the challenge of generating product recommendations for their online shop from historical sales data under these conditions.

Approach

To create product recommendations, a recommender system was developed using a deep learning model (LSTM) that learns to model purchase sequences based on historical shopping carts. This allows the current shopping cart information to be used during shopping to provide product recommendations to anonymous users. To solve the Cold Start Problem, products without sufficient history in the online shop were grouped and processed at an aggregated level. An API was developed to display recommendations in the online shop, allowing requests to be made to the recommendation engine from the shop. A serverless setup was deployed in the cloud to handle varying request volumes cost-effectively.

Result

The recommender model was developed and deployed on the client's cloud infrastructure. By using the developed recommender system, Feldschlösschen can display tailored product recommendations to customers in the online shop during their shopping experience. These product recommendations are shown at various points in the justDrink online shop and refined with each product added to the shopping cart.