Introducing our new metadataless algorithm : « Associated »

Say hi to content to content recommendation without metadata!

2 min. read

Associated is a type of content-to-content recommendation that works without metadata! It makes recommendations based on hidden user interaction patterns. However, this is no ordinary collaborative filtering, it’s more than that.

In general, collaborative filtering is a method of making automatic recommendations about a user’s interests by collecting preferences or taste data from many users. Essentially, it assumes that if people A and B both like a piece of content, person A will most likely like a piece of content that person B also likes. This explanation is a very simplistic way of defining collaborative filtering, as it is usually used in very large datasets allowing for less approximate conclusions. Nevertheless, it only takes into account popular content, arbitrarily recommending content that is likely to be consumed by users.

With Associated, we go beyond the simplistic concept of collaborative filtering. Based on user’s journey data and interaction patterns through the website, our algorithm is able to provide relevant recommendations. It’s about understanding users’ tastes and moods. In doing so, we avoid arbitrarily recommending the trendy content of the moment. Therefore, by taking the understanding process a step further, we are able to provide personalized recommendations without metadata.

But in concrete terms, how does it work?

In this image, we notice that in the data set, the most frequent pattern is reading the orange book after the blue one. We can then conclude that if a new user reads the blue book, chances are high that he’d want to read the orange one right after.

The main objective when developing this algorithm was to create an alternative recommendation recipe that can serve recommendation in 2 exceptional scenarios. Either the client does not have appropriate content metadata, but can integrate user interactions. Or having cold start problems for new users on the platform (note that this feature can work with a similar content recommendation).

Possible applications of such a feature could be to include it in search bars. It could be used for search optimization, displaying related content when the content you are looking for is not available. 

Find more projects on our roadmap.