The streamer has published a post on its Tech Blog that explains how it uses transfer learning, a form of machine learning, to find and analyse the connections that drive audiences to its many hundreds of original shows. This in turn drives their marketing, analytics and forward planning
In transfer learning the parameters learned from what is known as a source task improve the performance of a target task. In Netflix’s case, the overall source task is to find licensed content that most closely matches Netflix original content, as well as related viewer analytics.
A similarity map, where the machine learning model uses a show’s metadata, tags and summaries (what Netflix calls embeddings) to form links between titles. The marketing teams uses the results to categorise titles and quickly find a likely audience – core, curious and casual.
Another model compares the audience sizes of similar titles in a given country so that it knows where to spend its post production budget – on localisation services – and where additional localised marketing might be useful.
As well as saving time and money spent on time-consuming people-driven content curation, these self-supervised systems eliminate subjectivity and bias from the decision-making equation.