Off-the-shelf recommenders are a necessary building block for developing a personalised recommendation system, but they are not sufficient to solve personalisation problems by themselves. Such systems are designed to use either explicit or implicit data, but never both. They are unable to identify anomalous user activity, and cannot determine when one account is put to multiple distinct uses, for example combined personal and business purchases, or many users in a household sharing streaming media.

This talk will look at the pain points of recommendation algorithms and cover ways to overcome them in practise. We will dive into the example of data drift in recommendation using data from a music streaming service. A user’s behaviour is likely to change over time for multiple reasons: users might discover a new genre they like, they may associate negative memories with a song they used to love, or their taste may simply change as they grow up. Given a profile of a user’s tastes over time, it’s relatively straightforward to recommend content that will spark nostalgia.  However, “nostalgia” comes from Greek words meaning both “returning home” and “pain,” and the songs we loved once may not bring us joy today!

We’ll talk about how to identify changing tastes, find content which will feel like returning home without the pain, and give some simple suggestions for how to incorporate these findings into off-the-shelf recommenders to give a more robust user experience.