The Challenge:
Our client, a flight company, wanted an engine that would recommend ‘add-ons’ to new and returning customers making bookings, and even after the booking is made. There were a handful of challenges that came with this project: we had limited data (10 months) and over 800 million possible combinations of customers, flights and products.

Our Approach:
From the myriad of options available, we chose to use a collaborative filtering model approach based on matrix factorisation. This calculates a user’s product recommendations based on similar passengers’ preferences. Each passenger is given a score for all products, even if they have never previously purchased an item. Collaborative filtering is generally preferred where the breadth of information available is limited (i.e. the data is sparse), but user-level recommendations are required.

The Results:
Recommendation engines work well to increase sales and up-sell, and airlines can make a lot of money from ancillary sales. Bags, seating arrangements and schedule changes are often not included in the fare. By simplifying the initial solution recommended using aggregated data, we delivered an ancillary recommender model. In the future, our client knows which add-ons to package up together for the right customer.