AI Use Cases
Recommendation
Suggesting the items a user is likely to want next, based on behaviour and item data. It helps most when a catalogue is larger than anyone will browse.
Part of Cloud, Data & AI
Recommendation systems predict which items a user is likely to find relevant. The prediction can draw on the user's own behaviour, on what similar users did, or on the characteristics of the items themselves.
Where it helps
It pays off with catalogues too large to browse, such as retail, media, content libraries and marketplaces. If a user can reasonably see everything on offer, recommendation adds little. The effect is usually measured in engagement or conversion, and it comes from making relevant items easier to find, so it depends on the catalogue having them.
What it needs
It needs behavioural data in sufficient volume: what users viewed, bought or skipped. New users and new items have no history, which is known as the cold-start problem. A fallback, such as popular items in the same category or rules based on item attributes, is needed from the first release.
Failure modes to plan for
- Feedback loops, where recommending popular items makes them more popular and narrows what people see over time.
- Suggestions that fit the data and make no commercial sense, such as recommending a product the customer has just bought.
We design business rules around the model, such as excluding recent purchases and out-of-stock items, and test them as carefully as the model.