Sky is going beyond a simple recommendation engine for content across its vast library by asking customers what mood they are in and using machine learning to deliver the best results.
Speaking at the ReWork deep learning in retail and advertising summit in London, Jian Li, principal data scientist at Sky explained: "We have lots of content and a large database, so how do we use the latest tech to surface that to customers?"
By ingesting customer's viewing and device data Sky's team of data scientists can build machine learning models to make personalised recommendations. Now, Li's team has built a new model which looks to serve content according to a user's mood at the time, instead of simply what they have watched recently.
The first step was simply matching a customer's mood to a specific genre. "So if a customer is looking for something funny we can translate that to a genre in our database, namely: comedy," Li said.
However, he continued: "Not just comedy movies are funny, action movies can be funny, so we can limit ourselves to deliver content here. If someone wants something exciting, that translates to many genres. So we need to start to rank content by how funny it is, how exciting it is, so that we can break down genres. How?"
The issue here is one of classification, and the solution for Li was to turn to Sky's team of editors to tag content.
The training model consisted of a set of movies from Sky's catalogue to which they attributed a set of keywords, or tags, describing the nature of the content. The editorial team then assigned a mood label, for example action comedy, spy comedy, or family animation.
The data science team then built a model (Li would not be drawn into detailing what type) to learn the "semantic relationship between moods and key words with machine learning," Li said. The outcome is a weighted "set of semantic representations of moods," so a film like the spy comedy Johnny English would be assigned a score of 5.2 for 'funny', 3.8 for 'adventurous' and 1.4 for 'exciting', Li outlined in his presentation.
"We establish the similarity between the query and our content, rank it and deliver the most relevant to our customers," Li explained.
In practice this looks like a set of bubbles on the screen, each containing a mood keyword, such as 'exciting', 'scary', 'gritty', and 'feel-good'. How does it ascertain your mood? Well, it asks you.
Users can select one keyword to be delivered a set of movies related to this mood, or select multiple moods. Users can even pinch the bubbles to make them larger or smaller to weight them accordingly.
The results are a bit of a mixed bag for now. For a viewer in the mood for something funny and scary the model in Li's demo video delivered both Shaun of the Dead (OK) and Juno (what?).
Li did note that "different people have different definitions of funny" so his team is iterating on the model to improve it for cultural differences, for example. "This model gives us a good foundation to get semantic relationship between moods and these keywords," he said.
In terms of the rollout for this feature, Li is being coy, simply saying: "The only thing I can say is that the mood work will be powering Sky's product features in the near future."
Streaming platforms are having to compete for consumer's attention more than ever, as Netflix, Amazon and soon Disney will be looking to steal eyeballs away from traditional broadcasters like the BBC and Sky's platforms.
Netflix is the king of the recommendation engine in this space, and it is telling that it hasn't gone down this route yet, but the BBC is experimenting far and wide with ways to keep users of its iPlayer service engaged.
As reported by our sister title Techworld, Matthew Postgate, chief technology and product officer at the BBC said last year: "We know that online services which mould themselves to us as individuals, rather than reflect the institutions which provide them, is not a new concept, but we want to do it in a very BBC way. So not telling you what customers like you bought, but what you may love to watch or like to know."
Postgate also said that it was experimenting with what he calls "mood-based classification", where "you say how long you have got and say if you want to be happy, sad, excited or whatever and it runs algorithms across the archive and suggests shows on that basis." Sound familiar?
Read next: How the BBC rebuilt iPlayer on microservices