The goal was to set up a model that can make recommendations based on historic usage & content metadata. After A/B testing a couple of algorithms the choice was made to implement a hybrid model that retrains itself on a continuous basis as new usage data comes in. A lot of focus was put on making the model as efficient as possible in order to run very close to real-time.
As the output of the algorithm needs to be available in real-time, a pipeline was set up to write the recommendations to a DynamoDB database. An API gateway was then set up to be able to query this data quickly. In parallel a system of metrics was set up in order to be able to measure the efficiency of the recommendation model & the speed of the API call.
Datashift took the lead in model development and added additional architecture and engineering support where necessary. The collaboration with internal teams was key to ensure the predicting model generates results in an optimal environment. Development was done in an agile way in order to work in short, well-defined tasks and get quick user feedback.
Due to the close collaboration between Datashift and the in-house engineers, we achieved excellent results. Traffic is going up as well as interactions with users, giving us more data to further improve user experience. In the future we expect users to consume more content and stay longer on the platform. Bingewatch alert! :)
After this initial mission, the client asked us to create a roadmap to further use data and AI to improve user experience (e.g. cross brand recommendations).